A hybrid multi-scale indirect vision detector algorithm for heritage building fire detection Authors: Jian Liang, Junsheng Cheng Published: 2025 ABSTRACT AbstractAddressing the challenges of small target detection difficulties and unbalanced vision in indirect vision technology for heritage building fire detection, we propose the HMIV-DET hybrid multi-scale indirect vision detection algorithm. This algorithm incorporates three key innovations: Adaptive Indirect Vision Enhancement (AIVE) resolves vision imbalance issues in multi-mirror deployments through dynamic weight allocation; Adaptive Multi-kernel Feature Orchestration Block (AMFOBlock) employs parallel multi-scale feature extraction and gated activation mechanisms to enhance the capture capability for flame features of different sizes; and Hierarchical Cross-Scale Feature Fusion Network (HCSFPN) achieves comprehensive interaction of features across multiple semantic levels. Experiments on the self-constructed Heritage Building Indirect Vision Fire Dataset demonstrate that compared to the YOLO11 baseline, mAP50 and mAP50-95 improve by 2.8% and 3.3% respectively, significantly enhancing detection performance while maintaining lightweight characteristics.Subject terms: Engineering, Mathematics and computing Introduction Heritage buildings, as vital carriers of historical and cultural information, possess irreplaceable cultural value and historical significance. However, fire poses a severe threat to heritage structures, often resulting in irrecoverable losses to cultural heritage once it occurs1. Statistics indicate that the United Kingdom and Scotland experience hundreds of fire incidents affecting heritage buildings annually, with similar alarming loss figures worldwide. The predominant use of wooden structures in heritage buildings, combined with material deterioration due to age, significantly increases fire risk, urgently necessitating the establishment of effective early fire detection systems2. Timely and accurate fire detection not only minimizes heritage losses but also serves as a crucial technical means for protecting human cultural heritage and maintaining historical continuity. Traditional fire detection systems primarily rely on physical sensor devices such as smoke detectors and temperature sensors for fire identification. Smoke detectors detect smoke particles in the air through photoelectric or ionization principles, while temperature detectors monitor environmental temperature changes to determine fire occurrence3. Thermal detectors include fixed-temperature and rate-of-rise types, triggering alarms through set temperature thresholds and temperature rise rates respectively. However, traditional sensor methods suffer from delayed response, high false alarm rates, and sensitivity to environmental conditions4. Particularly in the complex indoor environments of heritage buildings, factors such as dust, humidity changes, and cooking activities easily lead to frequent false alarms, reducing system reliability and user trust. In recent years, with the rapid development of computer vision and deep learning technologies, image-based fire detection methods have gradually become research hotspots. Zhang et al. proposed a two-stage deep learning fire detection method for heritage buildings, achieving real-time fire identification through motion detection using Gaussian mixture models followed by the Fire-Det model5. Sathishkumar et al. introduced a forest fire and smoke detection method based on Learning without Forgetting, achieving 79.23% detection accuracy through transfer learning on pre-trained VGG16, Inception V3, and Xception models6. Singh et al. developed a unified wildfire detection system combining classification and segmentation, integrating custom convolutional neural networks with AlexNet for fire identification and employing UNet and SegNet architectures for precise burning area localization, achieving 88.19% accuracy in classification tasks7. Deep learning methods can automatically learn visual features of flames and smoke, demonstrating stronger environmental adaptability and lower false alarm rates compared to traditional approaches. Convolutional neural networks effectively extract flame color, texture, shape, and motion features, maintaining high detection accuracy under complex backgrounds and varying lighting conditions. Recently, heritage building fire detection algorithms based on object detection have matured progressively, leading us to adopt an object detection algorithm framework for research. Deep learning-based object detection algorithms have achieved significant progress in fire detection. Liang et al. proposed the MITA-YOLO method, combining indirect vision technology with YOLOv8, effectively addressing the limited camera deployment issues in heritage buildings8. Muhammad et al. developed a CNN fire detection system based on the GoogleNet architecture, achieving high-precision flame recognition in surveillance scenarios9. Multiple studies indicate that YOLO series algorithms have become the mainstream choice for real-time fire detection due to their favorable speed-accuracy balance. Saydirasulovich et al. implemented fire detection in smart city environments based on YOLOv6 algorithm, achieving 0.98 object recognition performance and 0.83 detection accuracy on a dataset containing 4000 images10. Geng et al. proposed the YOLOFM algorithm based on improved YOLOv5n network, integrating FocalNext and QAHARep-FPN structures, achieving high-precision detection on a fire and smoke dataset of 18644 images11. Kong et al. obtained 94% detection accuracy in coal mine fire detection based on improved YOLOv8s algorithm12. The newly released YOLO11 algorithm further optimizes network architecture, significantly reducing model parameters while maintaining high detection accuracy, with its adopted C3k2 module and C2PSA attention mechanism demonstrating outstanding performance in small target detection13. Given YOLO11’s superior performance and broad application prospects in fire detection tasks, this study selects YOLO11 as the baseline detection framework. Despite the mature performance of object detection and YOLO11 algorithms in fire detection tasks, numerous challenges remain in complex heritage building environments. First, the adoption of indirect vision results in size reduction of targets in mirror reflections, making small target flames difficult to detect accurately. Second, mirrors deployed at different positions form indirect vision regions of varying sizes due to distance differences from cameras, creating detection domain unfairness issues that affect overall system detection performance. Complex lighting variations, similar-colored interfering objects (such as sunsets, orange decorations), and smoke diffusion characteristics in complex environments also increase the risk of false positives and negatives. Therefore, we propose an indirect vision adaptive hybrid network heritage building fire detection algorithm named HMIV-DET, with the following main contributions: We propose the Adaptive Indirect Vision Enhancement module (AIVE), which systematically addresses vision imbalance issues caused by positional differences in multi-mirror deployments through dynamic weight allocation and spatial redistribution strategies. This module adaptively adjusts detection weights for different mirror regions, ensuring remote indirect vision receives attention comparable to proximal regions, achieving detection domain fairness. Comprehensive comparative experiments demonstrate that AIVE outperforms existing image enhancement methods including CLAHE, Retinex-based approaches, and deep learning-based LLNet by 1.4% in mAP50, validating its superiority in addressing mirror-specific spatial unfairness. We propose the Adaptive Indirect Vision Enhancement module (AIVE), which systematically addresses vision imbalance issues caused by positional differences in multi-mirror deployments through dynamic weight allocation and spatial redistribution strategies. This module adaptively adjusts detection weights for different mirror regions, ensuring remote indirect vision receives attention comparable to proximal regions, achieving detection domain fairness. Comprehensive comparative experiments demonstrate that AIVE outperforms existing image enhancement methods including CLAHE, Retinex-based approaches, and deep learning-based LLNet by 1.4% in mAP50, validating its superiority in addressing mirror-specific spatial unfairness. We design the Adaptive Multi-kernel Feature Orchestration Block (AMFOBlock), which differs from existing multi-scale feature extraction methods (e.g., MSBlock, ODConv) by integrating dynamic geometric convolution selection with gated activation mechanisms. Unlike static multi-scale approaches that apply fixed convolution patterns, AMFOBlock adaptively selects optimal geometric representations (square, horizontal strip, and vertical strip kernels) through attention-driven weight generation, specifically addressing the irregular morphological characteristics of flames in indirect vision where traditional fixed convolutions fail to capture directional texture variations caused by mirror reflections and perspective distortions. We design the Adaptive Multi-kernel Feature Orchestration Block (AMFOBlock), which differs from existing multi-scale feature extraction methods (e.g., MSBlock, ODConv) by integrating dynamic geometric convolution selection with gated activation mechanisms. Unlike static multi-scale approaches that apply fixed convolution patterns, AMFOBlock adaptively selects optimal geometric representations (square, horizontal strip, and vertical strip kernels) through attention-driven weight generation, specifically addressing the irregular morphological characteristics of flames in indirect vision where traditional fixed convolutions fail to capture directional texture variations caused by mirror reflections and perspective distortions. We propose the Hierarchical Cross-Scale Feature Fusion Network (HCSFPN), which transcends traditional feature pyramid networks (PAFPN, BIFPN) by introducing attention-driven adaptive fusion instead of simple summation or concatenation. While existing FPN variants employ fixed fusion weights or basic bidirectional pathways, HCSFPN dynamically allocates fusion weights based on global semantic context through the ADMFM module, enabling intelligent selection of critical features while suppressing background noise. This adaptive mechanism is particularly crucial for indirect vision scenarios where mirror regions contain varying signal-to-noise ratios due to distance-dependent resolution degradation. We propose the Hierarchical Cross-Scale Feature Fusion Network (HCSFPN), which transcends traditional feature pyramid networks (PAFPN, BIFPN) by introducing attention-driven adaptive fusion instead of simple summation or concatenation. While existing FPN variants employ fixed fusion weights or basic bidirectional pathways, HCSFPN dynamically allocates fusion weights based on global semantic context through the ADMFM module, enabling intelligent selection of critical features while suppressing background noise. This adaptive mechanism is particularly crucial for indirect vision scenarios where mirror regions contain varying signal-to-noise ratios due to distance-dependent resolution degradation. Related work YOLO11 baseline framework YOLO11, as the latest generation real-time object detection model released by Ultralytics, incorporates multiple innovative improvements based on the YOLO series algorithms. Compared to previous generation models, YOLO11 achieves higher mean average precision with reduced parameters, primarily attributable to its improved network architecture and optimization strategies. YOLO11’s network architecture consists of three core components: Backbone, Neck, and Detection Head. The backbone network adopts an improved CSPDarkNet architecture, replacing traditional C2f modules with C3k2 modules, achieving more efficient feature extraction through smaller convolutional kernels (33) and optimized cross-stage partial connections. The neck network employs an improved Feature Pyramid Network (FPN) structure, integrating Spatial Pyramid Pooling Fast (SPPF) modules and Path Aggregation Network (PANet) for effective multi-scale feature fusion. The detection head adopts a decoupled design, separating object classification and bounding box regression tasks into different branches, and incorporates the C2PSA (Cross-Scale Pixel Spatial Attention) mechanism to enhance cross-scale pixel spatial attention, significantly improving detection performance for small targets and complex scenes. Indirect vision enhancement technology In heritage building fire monitoring systems, indirect vision technology expands camera monitoring range through mirror reflection, effectively addressing vision limitations and equipment deployment constraints faced by traditional monitoring systems. However, existing indirect vision systems exhibit significant spatial unfairness issues, namely monitoring domain imbalance caused by distance differences in mirrors deployed at different positions. Related research in visual enhancement has proposed autoencoder-based deep learning methods to identify signal features in low-light images and adaptively enhance image brightness14, providing a theoretical foundation for addressing image quality issues in indirect vision. Furthermore, in occlusion processing technology, researchers have developed various methods to handle occlusion problems in object detection, including attention mechanism-based feature enhancement15 and multi-view feature fusion techniques16. Multi-scale object detection in complex scenarios While small object detection has been the primary focus in fire detection research, real-world heritage building environments present challenges across multiple object scales. Recent advances have addressed these challenges through architectural innovations in feature enhancement, attention mechanisms, and multi-scale aggregation. Li et al. proposed QGMN employing occlusion-guided multi-task learning to handle partial occlusions in UAV-based object detection37, demonstrating that explicit occlusion modeling significantly benefits detection tasks under complex scenarios. For fire detection in buildings with semi-occluded conditions, Shiratkar et al. developed a deep learning approach that effectively identifies flames despite partial visual obstruction38. Li et al. introduced SGINet with scale-robust complementary learning mechanisms for UAV imagery39, while Wu et al. improved YOLO11’s detection capability in dense smoke environments through enhanced multi-modal feature fusion40. Specialized small object detection methods have emerged to address the unique challenges of detecting targets with limited pixel representation. TPH-YOLOv541 replaces traditional detection heads with transformer prediction heads, leveraging self-attention mechanisms to enhance feature representation for small objects in aerial scenarios. DETR-based approaches have shown promise in small object detection through end-to-end transformer architectures. UAV-DETR42 proposes an efficient DETR framework tailored for UAV imagery, incorporating multi-scale feature fusion with frequency enhancement to capture both spatial and frequency information, along with frequency-focused downsampling to retain critical spatial details during feature extraction. SO-DETR43 addresses small object detection challenges through a dual-domain hybrid encoder that integrates spatial and frequency domains for effective multi-scale feature fusion, combined with an enhanced query selection mechanism using expanded IoU to dynamically optimize query initialization and knowledge distillation for efficient detection. Lightweight approaches like NanoDet-Plus44 demonstrate that anchor-free architectures with dynamic label assignment can achieve real-time performance on mobile devices while maintaining competitive accuracy. While these multi-scale and small object detection methods achieve impressive results on standard benchmarks, they assume spatially uniform image quality and conventional perspective geometry. However, indirect vision systems introduce unique challenges: mirrors at different distances create heterogeneous detection domains with varying pixel densities and perspective distortions, requiring explicit modeling of spatial unfairness inherent in multi-mirror deployments. Existing methods lack mirror-aware preprocessing mechanisms to establish detection equity across regions with distance-dependent resolution degradation. Related research on feature fusion networks Feature Pyramid Network (FPN) is an important technique for multi-scale feature fusion in object detection, constructing high-level semantic feature maps at all scales through top-down architecture and lateral connections, utilizing the inherent multi-scale pyramid hierarchy of deep convolutional networks to build feature pyramids with minimal additional computational cost17. The network design includes bottom-up and top-down pathways, where the bottom-up pathway is the conventional convolutional network feature extraction process with decreasing spatial resolution but enhanced semantic information as layers deepen; the top-down pathway reconstructs high-resolution semantically-rich features through upsampling and lateral connections18. To further improve feature fusion effectiveness, researchers have proposed various FPN improvement schemes: PANet adds bottom-up path aggregation based on FPN, shortening the information propagation path between bottom-layer and top-layer features19; Multi-scale Residual Aggregation Feature Pyramid Network (MSRA-FPN) employs unidirectional cross-layer residual modules, aggregating multi-layer features from bottom to top through triangular structures, further enhancing semantic information exchange between different levels20; addressing the semantic gap problem between features of different scales in traditional FPN, multi-scale semantic enhancement feature pyramid networks alleviate aliasing phenomena during feature fusion through semantic enhancement modules, semantic injection modules, and gated channel guidance modules21. These improvement methods all strive to enhance feature fusion effectiveness, providing important technical support for accurate object detection in complex scenes. Methods The HMIV-DET (Hybrid Multi-scale Indirect Vision Detection) algorithm proposed in this paper is an end-to-end detection framework specifically designed for heritage building fire detection. This algorithm addresses fire detection challenges in indirect vision environments through three core innovative modules working collaboratively. HMIV-DET first employs the Adaptive Indirect Vision Enhancement (AIVE) module as a preprocessing component, addressing vision imbalance issues in multi-mirror deployments through dynamic weight allocation and spatial redistribution strategies; subsequently adopts the Adaptive Multi-kernel Feature Orchestration Block (AMFOBlock) to replace traditional C3k2 modules, achieving precise capture of flame features of different sizes through dynamic kernel selection and gated activation mechanisms; finally employs the Hierarchical Cross-Scale Feature Integration Network (HCSFPN) to replace traditional PAFPN, significantly enhancing multi-scale feature representation capabilities through dense feature interaction and adaptive fusion mechanisms. The overall framework is shown in Fig. 1, with the three modules forming a complete detection pipeline from preprocessing to feature extraction to feature fusion. Overall architecture of the HMIV-DET framework. The system consists of three main stages: (1) Preprocessing stage where AIVE module balances spatial fairness across multiple mirror regions through adaptive enhancement; (2) Backbone feature extraction stage where AMFOBlock modules replace traditional C3k2 blocks to perform multi-scale adaptive feature extraction using parallel geometric convolutions and gated selection; (3) Neck fusion stage where HCSFPN replaces PAFPN to achieve hierarchical cross-scale feature integration through attention-driven multi-scale fusion modules (ADMFM). The framework processes 640640 input images and outputs multi-scale detection results for fire localization. Adaptive indirect vision enhancement (AIVE) Unlike traditional enhancement methods that apply uniform processing across entire images, indirect vision systems require mirror-aware preprocessing to address spatial unfairness caused by heterogeneous mirror positioning at different distances. We propose AIVE to explicitly model this unfairness through area-based adaptive weight calculation and region-specific enhancement, ensuring detection equity across all indirect vision domains. While the Target-Mask module can effectively isolate indirect vision regions and filter out non-target areas, mirror deployments at different spatial positions introduce significant detection domain imbalance. Specifically, mirrors deployed near cameras capture larger indirect vision regions with higher pixel density, while mirrors deployed at distant positions exhibit noticeably smaller coverage areas and reduced target resolution. This spatial unfairness creates an imbalanced detection environment where fire targets in remote mirrors may be underrepresented or completely missed during detection, thereby compromising the overall reliability and robustness of the fire monitoring system. To address this critical limitation, we propose the Adaptive Indirect Vision Enhancement (AIVE) module, serving as an intelligent preprocessing component designed to establish spatial fairness across all indirect vision regions. The AIVE module dynamically analyzes geometric features of each detected mirror region, calculates adaptive enhancement weights based on relative area differences, and applies targeted enhancement strategies to compensate for size-related detection disadvantages. Through sophisticated redistribution algorithms and visual enhancement techniques, AIVE ensures all indirect vision regions receive proportionally balanced attention in subsequent detection processes, effectively mitigating adverse effects of mirror placement constraints while maintaining fire detection capability throughout the entire monitoring domain, with its workflow shown in Fig. 2. Adaptive Indirect Vision Enhancement (AIVE) module processing pipeline. The workflow consists of four stages: (a) Automatic mirror detection using Hough Circle Transform to identify circular mirror regions and extract geometric parameters (center coordinates and radius); (b) Area-based adaptive weight calculation where smaller mirrors receive higher enhancement weights inversely proportional to their area to compensate for size disadvantages; (c) Multi-domain visual enhancement including brightness adjustment, contrast amplification, and sharpening applied to undersized mirrors; (d) Spatial redistribution algorithm that optimally repositions enhanced mirror regions to maximize detection coverage while minimizing inter-mirror interference. This comprehensive pipeline establishes spatial fairness across all indirect vision regions. As illustrated in Fig. 2, the AIVE processing pipeline systematically addresses spatial unfairness through four sequential stages: Stage (a) - Automatic Mirror Detection employs Hough Circle Transform to identify all circular mirror regions and extract geometric parameters including center coordinates and radius for each mirror , establishing the spatial distribution foundation. Stage (b) - Area-based Adaptive Weight Calculation computes enhancement weights inversely proportional to each mirror’s effective area, ensuring smaller distant mirrors receive higher compensation weights to address size disadvantages. Stage (c) - Multi-domain Visual Enhancement applies targeted brightness, contrast, and sharpening adjustments specifically to undersized mirrors with weights , with enhancement intensity proportional to computed weights. Stage (d) - Spatial Redistribution Algorithm optimally repositions enhanced mirror regions to maximize detection coverage while minimizing inter-mirror interference, establishing geometrically balanced layout for subsequent detection. The AIVE module employs a multi-stage processing pipeline, systematically addressing spatial imbalance issues through mathematical optimization and adaptive enhancement strategies. The workflow begins with automatic mirror detection using Hough Circle Transform, followed by area-based weight calculation, adaptive scaling, visual enhancement, and spatial redistribution. In the initial mirror detection phase, the system identifies circular mirror regions in the input image and extracts geometric parameters for each detected mirror. For a given mirror with center coordinates and radius , the effective area is calculated as: where represents the visibility factor, accounting for cases where partial occlusion or edge effects may reduce the effective monitoring area of mirror . The adaptive weight calculation mechanism establishes a fairness-based compensation system by computing enhancement weights inversely proportional to mirror size. The enhancement weight for mirror is determined by: where represents the average area of all N detected mirrors, and denotes a balancing factor controlling weight adjustment intensity. This formula ensures smaller mirrors receive proportionally higher enhancement weights while maintaining numerical stability through the offset term . Adaptive scaling transformation applies size compensation based on computed weights to achieve visual fairness across mirror regions. The scaling factor for each mirror is derived as: where represents the global scaling coefficient, maintaining overall image coherence while allowing individual mirror enhancement. The square root relationship ensures area-based scaling appropriately translates to linear dimension adjustment, achieving desired size normalization effects while preventing excessive distortion. For mirrors with enhancement weights exceeding 1 (), indicating undersized regions requiring enhancement, the AIVE module applies targeted visual enhancement through multi-domain processing. The enhanced pixel intensity is calculated as: where represents the original pixel intensity at coordinates (x, y) within mirror , serves as the reference intensity level, provides brightness enhancement, controls contrast amplification, represents convolution with sharpening kernel K when , with intensity coefficient proportional to enhancement requirements. The spatial redistribution algorithm redistributes processed mirror regions within the target image space to optimize detection coverage and minimize inter-mirror interference. The optimal placement coordinates for enhanced mirror are determined by: where T represents a transformation matrix incorporating both enhanced mirror dimensions and target image constraints, provides angular distribution for uniform spatial coverage, and denotes the centroid of available placement regions. This mathematical framework ensures optimal spatial utilization while maintaining geometric relationships between mirror regions and their corresponding monitoring areas. The implementation of the AIVE module significantly improves spatial unfairness inherent in multi-mirror indirect vision systems, establishing a more balanced and reliable foundation for heritage building fire detection. By dynamically compensating for size differences caused by distance and applying targeted enhancement strategies, AIVE effectively transforms heterogeneous collections of indirect vision regions into standardized, equitable detection domains where each monitoring area receives appropriate attention regardless of its original geometric constraints. This preprocessing enhancement not only improves detection sensitivity for remote or smaller mirror regions but also protects architectural integrity requirements necessitating indirect vision approaches while maintaining overall system efficiency, thereby advancing the technical frontier of heritage building fire monitoring systems. Adaptive multi-kernel feature orchestration block (AMFOBlock) Traditional C3k2 modules employ fixed 33 convolution kernels and lack dynamic feature selection mechanisms, limiting their ability to capture diverse flame morphologies in indirect vision. We propose AMFOBlock to achieve adaptive multi-scale feature extraction through dynamic kernel selection and gated activation, with its structure shown in Fig. 3. Architecture of the Adaptive Multi-kernel Feature Orchestration Block (AMFOBlock). The module employs dual-path parallel processing: (1) Parallel Scale-Adaptive Feature Extractor (PSAFE) path that splits input features and applies Multi-Geometric Adaptive Convolution (MGAC) with different kernel configurations to extract multi-scale features; (2) Gated Selection Activation Unit (GSAU) path that generates adaptive gating signals through depthwise separable convolutions and sigmoid activation to selectively enhance critical features. Both paths integrate through residual connections with batch normalization for stable training convergence. AMFOBlock employs a dual-path parallel processing architecture, achieving efficient multi-scale feature extraction through hierarchical normalization, dynamic feature mixing, and gated activation mechanisms. The module’s overall feature transformation process follows a residual learning paradigm, with its workflow expressed as: where represents the input feature tensor, and respectively denote the Parallel Scale-Adaptive Feature Extractor (PSAFE) and Gated Selection Activation Unit (GSAU), is the batch normalization operation. Hierarchical scaling parameters achieve channel-level adaptive adjustment through broadcasting mechanism , ensuring stable convergence of different feature channels during training. The Parallel Scale-Adaptive Feature Extractor (PSAFE) achieves adaptive multi-scale feature extraction through parallel multi-kernel processing, with its feature fusion mechanism defined as: where , are feature subsets after equal channel dimension splitting, and represent Multi-Geometric Adaptive Convolution (MGAC) operations with different kernel sizes, is the 11 convolution feature fusion function. This design enables different scale convolution branches to focus on learning scale-specific feature representations through feature channel grouping and parallel processing, ultimately achieving effective multi-scale feature fusion through point convolution. The Gated Selection Activation Unit (GSAU) employs gating mechanisms to achieve selective feature activation and nonlinear transformation, with its computation process expressed as: where , , , represents 33 depthwise separable convolution operation, denotes element-wise multiplication (Hadamard product). This gating mechanism modulates feature activation in the U branch through attention weights generated by the V branch, achieving adaptive feature selection and information flow control,with its structure shown in Fig. 4. Multi-Geometric Adaptive Convolution (MGAC) module structure. The module performs parallel processing using three complementary convolution operators: (1) Square convolution kernel (kk) for capturing local neighborhood features; (2) Horizontal strip convolution (1k_band) for extracting horizontal directional texture; (3) Vertical strip convolution (k_band1) for vertical directional features. Dynamic weight generation through global average pooling and 11 convolution produces adaptive attention weights for each branch, with softmax normalization ensuring proper weight distribution. The weighted fusion of multi-branch outputs, followed by batch normalization and SiLU activation, enables adaptive selection of optimal geometric representations for diverse flame morphologies. The Multi-Geometric Adaptive Convolution (MGAC) module achieves adaptive extraction of flame target features at different orientations and scales through parallel processing of multi-geometric shape convolution kernels and dynamic weight allocation mechanisms. The module designs three complementary convolution operators: square convolution kernel for capturing local neighborhood features, horizontal strip convolution kernel specifically extracting horizontal directional texture information, and vertical strip convolution kernel focusing on vertical directional structural features. The dynamic weight generation mechanism calculates adaptive weights for each convolution branch through global context awareness and channel attention: where represents global adaptive average pooling operation, is the 11 convolution layer for weight generation with output channels of 3C, operation reshapes the weight tensor to five-dimensional tensor [3, B, C, 1, 1] for subsequent branch weighted computation. The Softmax operation ensures probabilistic distribution properties of the three branch weights, i.e., . The weighted fusion process of multi-branch dynamic convolution is defined as: where , , respectively represent square , horizontal strip , and vertical strip depthwise separable convolution operations, is the dynamic weight tensor for the i-th branch, is the SiLU activation function. This design achieves dynamic selection of convolution kernels with different geometric shapes through soft attention mechanisms, enabling the model to adaptively adjust feature extraction strategies based on spatial distribution characteristics of input features. The AMFOBlock module significantly enhances YOLO11’s feature representation capability and detection accuracy in heritage building fire detection tasks through integrating adaptive multi-scale feature extraction, dynamic weight allocation, and gated activation mechanisms. This module effectively overcomes the static feature extraction limitations of traditional C3k2 structures, achieving parallel adaptive extraction of flame features of different sizes through PSAFE, while GSAU enhances the model’s sensitivity to critical flame features through selective feature activation mechanisms. Particularly in complex indirect vision environments, MGAC’s multi-directional convolution design can more accurately capture irregular geometric morphological features of flames, with dynamic weight mechanisms adaptively adjusting contribution degrees of different convolution branches based on scene characteristics, thereby significantly improving small target flame detection accuracy and overall model generalization capability while maintaining computational efficiency. Hierarchical cross-scale feature fusion network (HCSFPN) Traditional PAFPN’s linear feature aggregation cannot effectively balance multi-scale feature importance, potentially masking critical flame features during fusion. We propose HCSFPN to achieve adaptive fusion through attention-driven multi-scale fusion modules and hierarchical aggregation strategies, with its structure shown in Fig. 5(a). (a) Architecture of Hierarchical Cross-Scale Feature Fusion Network (HCSFPN). The network constructs dense multi-scale feature interaction paths where each scale level receives inputs from multiple semantic levels through ADMFM modules, enabling comprehensive cross-level feature complementarity. Features flow through progressive fusion stages (P3P4P5 top-down path and P5P4P3 bottom-up path) with rich lateral connections. (b) Attention-Driven Multi-scale Fusion Module (ADMFM) structure showing channel unification processing, global average pooling for context extraction, MLP-based attention weight generation via softmax, and weighted multi-scale feature summation for intelligent adaptive fusion. The HCSFPN architecture constructs a dense multi-scale feature interaction network, achieving comprehensive fusion and complementarity of features at different semantic levels. The architecture adopts a hierarchical progressive feature aggregation strategy, where each scale level receives feature inputs from multiple different levels through ADMFM modules, forming rich feature fusion paths. The overall feature flow of the network can be expressed as a multi-level progressive fusion process: where represents the i-th scale feature map after t-th level processing, is the adjacent scale set associated with the i-th scale, is the attention-driven multi-scale fusion module operation, is the original feature from the backbone network. This design enables features at each scale to obtain information supplementation from multiple semantic levels through cross-level feature connections, thereby enhancing the network’s perception capability for flame targets of different sizes. HCSFPN’s dense connection mechanism further achieves efficient semantic information flow through multi-path feature propagation, with its global feature fusion process modeled as: where L represents the number of output feature levels, is the global aggregation function for the i-th layer, is the number of ADMFM processing stages for the i-th layer, denotes serialized feature processing operations, is the input feature set for the k-th stage of the i-th layer. The Attention-Driven Multi-scale Fusion Module (ADMFM) achieves intelligent fusion of multi-scale features through adaptive attention mechanisms, with its core idea being to dynamically allocate fusion weights for different input features based on global semantic information of features,with its structure shown in Fig. 5(b). The module first maps heterogeneous features from different levels to a unified feature space through channel unification processing, then generates adaptive attention weights based on global context information. The feature unification and weight generation process of ADMFM can be expressed as: where represents the i-th input feature map, is the channel dimension unification function (11 convolution when , otherwise identity mapping), is the global adaptive average pooling operation, is the multi-layer perceptron attention generation network, is the generated multi-scale attention weight tensor. This weight generation mechanism can adaptively evaluate the contribution of different input features to the final fusion result through perception of global semantic information, ensuring important features receive higher weight allocation. The final feature fusion process of ADMFM achieves effective integration of multi-scale features through weighted summation, with its mathematical expression as: where represents element-wise multiplication operation, function expands the unified feature tensor to five-dimensional form to match attention weight dimensions, is the final fused feature map. Through attention-weighted approach, this fusion mechanism enables feature branches with stronger semantic expression capability to dominate in the final fusion result, while feature branches with more noise or weaker semantic information are automatically suppressed, thereby achieving intelligent fusion and optimal selection of multi-scale features. The HCSFPN feature fusion network architecture significantly enhances YOLO11’s multi-scale feature representation and fusion capability in heritage building fire detection tasks by employing attention-driven multi-scale fusion modules and dense connection mechanisms. This architecture effectively addresses the issues of uneven weight distribution and limited semantic information propagation in traditional PAFPN during feature fusion. Through the adaptive attention mechanism of ADMFM modules, intelligent fusion of features at different scales is achieved, enabling the model to more accurately capture flame target features of different sizes. Particularly in indirect vision environments, HCSFPN’s multi-level aggregation strategy can fully utilize shallow detail information and deep semantic information, achieving cross-scale information complementarity through dense feature interaction paths, thereby significantly improving small target flame detection accuracy and the model’s adaptability to complex scenes while maintaining detection capability for large-size flame targets. Experimental results and analysis Self-constructed dataset This study selected the Former Site of the Military Committee of the Guangdong District Committee of the Communist Party of China, located in Guangzhou City, Guangdong Province, China, as the research object. As a provincial-level cultural relics protection unit in Guangdong Province, this former site bears irreplaceable cultural and historical value. However, its complex spatial structure and the requirement to preserve historical features render conventional monitoring technologies based on direct vision inadequate, resulting in prominent demands and significant challenges for fire prevention monitoring. Given the uniqueness of heritage buildings, real fire tests pose severe safety risks. To address this constraint, high-fidelity virtual fire scenes were constructed using 3DMAX software for subsequent fire detection experiments, with the 3D modeling process based on on-site survey data of the former site. During modeling, the actual structural components (e.g., wooden beams, walls, and historical relic display areas) and environmental parameters (e.g., natural light intensity and indoor occlusion conditions) of the building were strictly matched to ensure high scene reproducibility. To demonstrate the advantage of indirect vision technology in expanding the monitoring range of cameras, four areas were designated as target monitoring regions, including blind spots that are inaccessible under direct vision. Four wide-angle mirrors were deployed without damaging the original structure of the building, and the angles of the mirrors were adjusted to ensure that the indirect vision range of each mirror fully covered its corresponding target monitoring region. This setup simulated the coverage effect of indirect vision on occluded high-risk areas in real scenarios. Subsequently, a virtual camera was deployed in the lobby area of the 3D model, with its lens facing the main entrance of the building, to capture fire scene images containing indirect vision information. All captured images were preprocessed using the Target-Mask module proposed in our previous research, which filters out non-indirect vision regions through pixel-level masking. The finally constructed fire detection dataset, Heritage Building Indirect Vision Fire Dataset (Heritage-IVF), contains 1000 mask-processed fire images saved in JPG format with pixel dimensions of 16452493. The dataset was randomly divided into three subsets at a ratio of 7:1.5:1.5, namely a training set (700 images), a validation set (150 images), and a test set (150 images), and the category is “fire”. The construction workflow of this dataset is illustrated in Fig. 6. Construction workflow of the Heritage Building Indirect Vision Fire Dataset (Heritage-IVF). The process consists of five stages: (1) On-site surveying and measurement of the Guangdong Military Committee Former Site to collect structural parameters; (2) High-fidelity 3D modeling using 3DMAX software with accurate reproduction of architectural elements, materials, and lighting conditions; (3) Strategic mirror deployment in the virtual environment (4 wide-angle mirrors) to cover monitoring blind spots while preserving architectural integrity; (4) Virtual fire scene generation at various locations with realistic flame rendering and smoke simulation; (5) Image capture from virtual camera followed by Target-Mask preprocessing to isolate indirect vision regions. Final dataset contains 1000 images (16452493 pixels) divided into training (700), validation (150), and test sets (150). Experimental environment and parameter settings The experimental environment of this study is based on Ubuntu 20.04 operating system, using PyTorch 1.13.1 deep learning framework with Python 3.8 for model development and training. Hardware configuration includes Intel Xeon(R) CPU E5-2682 v4 @ 2.50GHz processor, NVIDIA GeForce RTX 3090 graphics card (24GB video memory), 32GB memory, and 50GB hard disk storage space (expanded by 60GB), equipped with CUDA 11.7 computing platform providing powerful computational resource support for deep learning model training and inference. Model training hyperparameter settings are as follows: input image size uniformly adjusted to 640640 pixels to accommodate network input requirements, batch size set to 64, SGD stochastic gradient descent optimizer used for parameter optimization with momentum parameter of 0.937 and weight decay coefficient of 0.0005 to prevent overfitting. Training epochs set to 100 to ensure sufficient model convergence. Learning rate adopts cosine annealing schedule with initial value of 0.01, minimum value of 0.0001, and warm-up period of 3 epochs. Data augmentation strategies include mosaic augmentation (probability 1.0), random horizontal flip (probability 0.5), random scaling (range 0.5-1.5), and HSV color space augmentation (hue shift ±0.015, saturation shift ±0.7, value shift ±0.4). Model selection criterion uses validation set mAP50 metric, with the best-performing checkpoint saved for final evaluation. To ensure experimental reproducibility and fairness, all comparative and ablation experiments employ identical random seed (42), data augmentation configuration, and training hyperparameters, with only architectural components varying as specified. All other training parameters adopt YOLO11 model default settings and use officially provided pre-trained weight files for initialization to enhance model generalization capability. Evaluation metrics This study employs standard evaluation metrics from the object detection field to comprehensively evaluate model performance. Precision (P) measures the proportion of actual positive samples among samples predicted as positive by the model, reflecting detection result accuracy; Recall (R) represents the proportion correctly detected among actual positive samples, embodying the model’s target detection capability. For comprehensive evaluation following COCO dataset standards, we employ multiple Average Precision (AP) metrics: mAP50 represents average precision at IoU threshold of 0.5, mAP50-95 represents average precision across IoU thresholds from 0.5 to 0.95 with 0.05 increments, providing a more stringent evaluation of localization accuracy. Additionally, we report scale-specific metrics including APS (small objects with area < 32), APM (medium objects with 32 < area < 96), and APL (large objects with area > 96), which are particularly important for evaluating multi-scale detection capability in indirect vision fire detection scenarios where flame targets appear at varying scales depending on mirror distance and fire development stage. Additionally, computational complexity metric GFLOPs (Giga Floating-point Operations) is used to evaluate model computation amount, Parameters (Params) measures model scale size, and frame rate (FPS, Frames Per Second) evaluates model real-time processing capability. These metrics comprehensively reflect fire detection model comprehensive performance from multiple dimensions including detection accuracy, model efficiency, and inference speed, providing important reference basis for practical deployment of heritage building fire monitoring systems. Ablation experiments Ablation experiment on AMFOBlock (All following experiments use AIVE algorithm processed baseline) To verify the effectiveness of our proposed AMFOBlock, we conducted ablation experiments on different components within AMFOBlock, focusing on exploring the impact of different convolution kernel configurations on fire detection performance. Experimental results shown in Table 1 systematically evaluate six configuration schemes including traditional square convolution kernels, asymmetric strip convolution kernels, and multi-kernel mixing strategies. Table 1Ablation experiment on convolution kernel configurations in AMFOBlock (All following experiments processed with AIVE algorithm).Kernel ConfigurationK_squareK_bandP(%)R(%)mAP50(%)mAP50-95(%)APS(%)APM(%)1. base--92.585.891.562.656.865.0233-92.585.692.062.456.564.8355-92.383.991.762.156.264.54-17, 7192.784.191.962.456.664.95-111, 11192.385.792.163.157.265.46. ours33111, 11193.187.292.562.857.065.2 Ablation experiment on convolution kernel configurations in AMFOBlock (All following experiments processed with AIVE algorithm). Experimental results indicate that single convolution kernel configurations exhibit obvious performance bias across different evaluation metrics: square convolution kernels (33, 55) maintain stable precision but relatively low recall; asymmetric strip convolution kernels demonstrate certain advantages in capturing directional features, particularly the (111, 111) configuration achieving 63.1% mAP50-95, improving by 0.5% over baseline. Notably, our proposed multi-kernel mixing strategy (33 + 111/111) achieved optimal performance across all evaluation metrics, with mAP50 and mAP50-95 improving by 1.0% and 0.2% respectively compared to baseline. This significant performance improvement fully validates that AMFOBlock can more comprehensively capture spatial features and directional texture information of flame targets through dynamically fusing multi-scale geometric shape convolution kernels, thereby achieving balanced optimization of detection accuracy and recall. Comparative experiment on AIVE with existing image enhancement methods To validate the superiority of our proposed AIVE module, we conducted comprehensive comparative experiments with existing state-of-the-art image enhancement methods. We selected five representative approaches covering different enhancement paradigms: (1) CLAHE (Contrast Limited Adaptive Histogram Equalization) for traditional adaptive enhancement; (2) Retinex-based enhancement for low-light image improvement; (3) Deep learning-based LLNet for learned enhancement; (4) Multi-Scale Retinex with Color Restoration (MSRCR) for color-preserved enhancement; and (5) Gamma correction with adaptive parameter selection. All methods were applied as preprocessing steps before detection, maintaining identical training configurations for fair comparison. Table 2 presents comprehensive performance comparison across computational efficiency and detection accuracy metrics. Table 2Comparison of AIVE with existing image enhancement methods.Enhancement MethodGFLOPsParams(M)P(%)R(%)mAP50(%)mAP50-95(%)APS(%)APM(%)No Enhancement6.32.5892.184.290.160.554.263.1CLAHE6.32.5892.384.890.661.154.963.6Retinex-based6.32.5892.085.190.460.954.763.4LLNet146.32.5892.685.391.061.555.364.0MSRCR6.32.5892.484.990.861.355.163.8Adaptive Gamma6.32.5892.285.090.561.054.863.5AIVE (Ours)6.32.5892.585.891.562.656.865.0 Comparison of AIVE with existing image enhancement methods. The experimental results demonstrate AIVE’s substantial advantages over existing enhancement methods. Traditional CLAHE achieves only 0.5% mAP50 improvement over no enhancement, limited by its global histogram manipulation that fails to address the fundamental spatial unfairness issue in multi-mirror indirect vision systems. Retinex-based methods show minimal improvement of 0.3% in mAP50, as they focus on illumination-reflectance decomposition suitable for general low-light scenarios but lack the mirror-specific spatial awareness required for indirect vision environments. Deep learning-based LLNet demonstrates better performance with 0.9% mAP50 improvement, validating the benefit of learned enhancement strategies. However, LLNet’s generic low-light enhancement paradigm cannot explicitly model the geometric size disparities between mirrors at different distances. MSRCR and Adaptive Gamma show moderate improvements of 0.7% and 0.4% in mAP50 respectively but remain constrained by their uniform processing approach that treats all image regions equally, ignoring the heterogeneous nature of multiple mirror regions. In contrast, AIVE achieves the highest performance across all metrics: 92.5% precision, 85.8% recall, 91.5% mAP50, and 62.6% mAP50-95, representing improvements of 1.4% and 2.1% in mAP50 and mAP50-95 over baseline. The superior performance stems from AIVE’s three unique advantages: (1) Mirror-aware spatial fairness through area-based adaptive weight calculation that explicitly compensates for distance-induced size disparities; (2) Region-specific enhancement that applies differential processing intensities based on computed weights, ensuring remote small mirrors receive proportionally stronger enhancement; (3) Geometric redistribution that optimally repositions mirror regions to minimize interference while maximizing detection coverage. These mirror-specific design principles, absent in general image enhancement methods, enable AIVE to fundamentally address the spatial unfairness challenge in indirect vision fire detection, establishing a balanced detection foundation that benefits all subsequent processing stages. Comparative ablation experiment on different feature fusion strategies To verify the effectiveness of our proposed ADMFM, we conducted comparative ablation experiments on different feature fusion strategies, systematically evaluating ADMFM’s contribution of adaptive attention mechanisms to multi-scale feature fusion. Experimental results shown in Table 2 comprehensively compare ADMFM with four traditional feature fusion methods including direct summation Element-wise Add, maximum value selection based MaxPool, channel concatenation Concatenation, and fixed weight Weighted Sum strategy. Experimental results show significant differences in performance among different fusion strategies: although MaxPool achieves 92.9% precision, its recall is only 84.3%, indicating maximum pooling strategy easily loses fine-grained feature information; Concatenation method’s mAP50-95 decreases by 1.7% compared to baseline due to feature redundancy from simple concatenation; fixed weight fusion strategy (Weighted Sum) shows improved recall of 86.1% but lacks adaptability limiting overall performance improvement. In contrast, our proposed attention mechanism-based ADMFM achieved optimal performance across all evaluation metrics, with recall improving by 1.1% compared to Element-wise Add, while mAP50 and mAP50-95 improve by 0.7% and 0.4% respectively compared to baseline. This comprehensive performance improvement fully validates that ADMFM can more precisely retain critical flame feature information while effectively suppressing background noise interference through adaptive attention weights dynamically adjusting fusion proportions of features at different scales, thereby achieving intelligent fusion and optimal selection of multi-scale features. Overall ablation experiment To verify the effectiveness of our proposed HMIV-DET, we conducted systematic ablation experiments, evaluating each core component’s contribution to overall performance by progressively integrating them. Experimental results shown in Table 3 use YOLO11 as baseline, sequentially employing AIVE, AMFOBlock, and HCSFPN three innovative modules, comprehensively evaluating each component’s performance contribution and computational efficiency. Table 3Comparative ablation experiment on different feature fusion strategies (All following experiments processed with AIVE algorithm).Fusion StrategyDescriptionP(%)R(%)mAP50(%)mAP50-95(%)APS(%)APM(%)Element-wise AddDirect summation92.585.891.562.656.865.0MaxPoolMaximum pooling92.984.392.062.256.464.6ConcatenationChannel concatenation92.585.491.160.955.163.3Weighted Sum (Fixed)Fixed weights fusion92.486.191.762.056.264.4ADMFM (ours)Attention-based adaptive92.886.992.263.057.165.4 Comparative ablation experiment on different feature fusion strategies (All following experiments processed with AIVE algorithm). Experimental results fully demonstrate the effectiveness of each innovative component in addressing specific technical challenges. First, after employing the AIVE module (Model 2), recall improves by 1.6% and mAP50-95 improves by 2.1%, validating that AIVE effectively addresses vision imbalance issues in multi-mirror deployments through dynamic weight allocation and spatial redistribution strategies, significantly enhancing flame target detection capability in remote small-size mirror regions. Second, after integrating the AMFOBlock module (Model 3), mAP50 improves by 1.0% and recall substantially improves by 1.4%, while achieving model lightweighting, fully proving that dynamic multi-kernel mixing mechanisms successfully address limitations of traditional fixed convolution unable to effectively capture diverse flame morphological features through adaptively selecting convolution kernels of different geometric shapes, improving feature expression capability while maintaining computational efficiency. When employing the HCSFPN module alone (Model 4), although GFLOPs increase by 0.6, mAP50 and mAP50-95 further improve by 0.7% and 0.4% respectively, validating that this module effectively overcomes critical feature loss issues caused by linear fusion in traditional PAFPN through adaptive attention weights and dense feature interaction paths, achieving intelligent aggregation of multi-scale semantic information. Finally, the complete HMIV-DET architecture achieved comprehensive performance breakthroughs with only 0.3 GFLOPs computational overhead increase and 0.07M parameter reduction: precision improves by 0.9%, recall substantially improves by 3.6%, mAP50 improves by 2.8%, and mAP50-95 improves by 3.3%, fully validating that the synergistic effect of three innovative modules successfully constructs an efficient and robust heritage building indirect vision fire detection framework. Comparative experiments Comparative experiment on different C3k2 modules To verify the effectiveness of our proposed AMFOBlock, we conducted comparative experiments on different feature pyramid network improvement methods, systematically evaluating AMFOBlock’s performance advantages relative to existing advanced feature extraction modules. Experimental results shown in Table 4 comprehensively compare AMFOBlock with five representative C3k2 improvement schemes including MSBlock, PConv, RepNCSPELAN, ODConv, and C3k2-Faster modules, evaluating each method’s performance from six dimensions. Table 4Overall ablation experiment.ModelAIVEAMFOBlockHCSFPNGFLOPsParams(M)P(%)R(%)mAP50(%)mAP50-95(%)APS(%)1. base6.32.5892.184.290.160.554.226.32.5892.585.891.562.656.835.92.3593.187.292.562.857.046.92.6392.886.992.263.057.15. ours6.62.5193.087.892.963.858.5 Overall ablation experiment. Experimental results indicate significant differences in performance and efficiency trade-offs among different improvement strategies. Although RepNCSPELAN achieved the lowest computational complexity, precision was only 91.9%; PConv excelled in parameter count but recall dropped to 84.6%, affecting overall detection performance; C3k2-Faster achieved the highest precision of 93.3% but with relatively high computational overhead and parameters. In contrast, our proposed AMFOBlock achieved optimal performance-efficiency balance: in computational complexity, GFLOPs reduce by 6.3% and parameters reduce by 8.9%, achieving effective model lightweighting; in detection performance, precision improves by 0.6%, recall substantially improves by 1.4%, mAP50 improves by 1.0%, and mAP50-95 improves by 0.2%. This comprehensive performance advantage fully validates that AMFOBlock significantly enhances feature extraction capability while maintaining lightweight characteristics through dynamic multi-kernel mixing mechanisms and gated activation strategies, providing a more efficient and robust solution for heritage building fire detection. Comparative experiment on different feature pyramid networks To verify the effectiveness of the proposed HCSFPN (Hierarchical Cross-Scale Feature Fusion Network), we conducted comparative experiments on different feature pyramid network architectures. Experimental results shown in Table 5 selected PAFPN as the baseline method and comprehensively compared performance with mainstream feature pyramid improvement methods including HSFPN, GFPN, MAFPN, and BIFPN. Table 6Comparative experiment on different feature pyramid networks (All following experiments processed with AIVE algorithm).ModelGFLOPsParams(M)P(%)R(%)mAP50(%)mAP50-95(%)APS(%)APM(%)PAFPN6.32.5892.585.891.562.656.865.0HSFPN265.71.8592.683.791.261.655.864.0GFPN278.43.6992.586.492.062.957.065.3MAFPN287.22.7392.985.592.162.556.764.9BIFPN296.41.9591.585.091.762.456.564.8HCSFPN (ours)6.92.6392.886.992.263.057.165.4 Comparative experiment on different feature pyramid networks (All following experiments processed with AIVE algorithm). Table 5Comparative experiment on different C3k2 modules (All following experiments processed with AIVE algorithm).ModelGFLOPsParams(M)P(%)R(%)mAP50(%)mAP50-95(%)APS(%)APM(%)C3k2 (base)6.32.5892.585.891.562.656.865.0C3k2-MSBlock226.72.5592.487.091.762.656.965.0C3k2-PConv235.82.2392.884.690.761.155.363.5RepNCSPELAN245.22.1691.985.991.362.256.464.6C3k2-ODConv255.72.6792.085.090.961.455.663.8C3k2-Faster236.12.3393.386.292.462.356.564.7AMFOBlock (ours)5.92.3593.187.292.562.857.065.2 Comparative experiment on different C3k2 modules (All following experiments processed with AIVE algorithm). Experimental results demonstrate that among all comparison methods, HCSFPN exhibited excellent performance across various evaluation metrics. From computational efficiency perspective, HCSFPN’s GFLOPs of 6.9, although higher than HSFPN and PAFPN, was significantly lower than GFPN and MAFPN, maintaining moderate computational complexity. In detection performance, HCSFPN achieved 92.8% precision and 86.9% recall, with recall being the highest among all methods, improving by 0.5% over the second-best GFPN and improving by 1.1% over baseline PAFPN. More importantly, HCSFPN achieved best results in both key metrics mAP50 and mAP50-95, reaching 92.2% and 63.0% respectively, with mAP50-95 improving by 0.1% over second-best GFPN and improving by 0.4% over PAFPN baseline. These quantitative results fully demonstrate that HCSFPN achieves comprehensive performance improvement while maintaining reasonable computational overhead through employing hierarchical cross-scale feature fusion and multi-level aggregation strategies, validating the method’s effectiveness and superiority. Comparative experiment with different mainstream SOTA models Figure 7 compares detection results across four representative scenarios, revealing the advantages of our algorithm in tackling core challenges of indirect-view fire detection. Quantitative analysis shows that in scenarios (a) and (b), for small-scale flame targets in the upper-right mirror region, YOLO11 achieves detection confidence of 0.82 and 0.87 respectively, while HMIV-DET reaches 0.85 and 0.90. This performance gain validates the spatial equity compensation mechanism of the AIVE module: through adaptive weight calculation based on area differences and visual enhancement strategies, it alleviates the inadequate feature representation caused by the relatively smaller area of distal mirror surfaces, allowing remote viewing regions to achieve detection sensitivity equivalent to proximal areas. Qualitative comparison of detection results between YOLO11 baseline and HMIV-DET across four representative scenarios. Scenarios (a-b) demonstrate improved detection confidence for small-scale flame targets in the upper-right mirror region, where HMIV-DET achieves 0.85 and 0.90 confidence compared to YOLO11’s 0.82 and 0.87, validating AIVE’s spatial equity compensation mechanism. Scenarios (c-d) showcase superior bounding box quality and flame contour fitting, attributed to AMFOBlock’s multi-geometric adaptive convolution capturing directional flame features and HCSFPN’s hierarchical semantic fusion preserving fine-grained details. HMIV-DET maintains high confidence outputs across varying flame scales and spatial locations, demonstrating robust multi-scale detection capability. The bounding box quality in scenarios (c) and (d) reveals the feature refinement capability of the AMFOBlock module. Our method achieves 0.87 confidence in scenario (c), with flame contour fitting superior to baseline models. This is attributed to MGAC’s dynamic fusion of square and strip convolutional kernel features, adaptively capturing morphological characteristics of flames in different directions, while the GSAU gating mechanism effectively suppresses background noise. HMIV-DET maintains high confidence outputs for both small and medium-to-large targets. This cross-scale robustness demonstrates the multi-level semantic aggregation advantages of HCSFPN: through ADMFM attention-driven fusion and dense connection strategies, it enables comprehensive interaction between shallow detail features and deep semantic features, effectively addressing the small target feature submersion problem in traditional PAFPN linear fusion, establishing a robust and precise multi-scale feature representation system under complex indirect-view environments. To comprehensively validate the effectiveness of HMIV-DET, we conducted extensive comparative experiments with three categories of state-of-the-art object detection models: general YOLO series models, improved YOLO variants, and small object detection specialized methods. Table 7 presents the comprehensive performance comparison across all evaluated models on the Heritage-IVF dataset. Table 7Comprehensive comparison with mainstream SOTA models.ModelCategoryGFLOPsParams(M)P(%)R(%)mAP50(%)mAP50-95(%)APS(%)APM(%)General YOLO SeriesYOLOv5n30YOLO7.12.590.382.688.258.752.361.5YOLOv8n31YOLO8.13.091.984.490.560.954.863.7YOLOv9t24YOLO7.72.090.882.288.158.351.960.8YOLOv10n32YOLO6.52.389.981.587.657.851.260.1YOLOv11n13YOLO6.32.692.184.290.160.554.263.1YOLOv12n33YOLO5.82.591.183.889.459.953.562.4YOLOv13n34YOLO6.12.490.482.988.758.652.161.0Improved YOLO Variantshyper-yolo-t35Hypergraph9.53.690.783.389.259.853.262.2mamba-yolo36State-space13.25.891.283.789.659.753.162.1Small Object Detection Specialized MethodsTPH-YOLOv541Transformer237.386.791.284.190.361.055.864.5UAV-DETR42DETR-based43.113.190.883.589.560.255.563.2SO-DETR43DETR-based33.312.191.384.290.261.356.864.0NanoDet-Plus44Anchor-free1.51.1788.981.287.857.552.361.5HMIV-DET (Ours)Hybrid Multi-scale6.62.5193.087.892.963.858.566.2 Comprehensive comparison with mainstream SOTA models. Table 7 demonstrates HMIV-DET’s superior performance across all comparison categories. HMIV-DET achieves 92.9% mAP50 and 63.8% mAP50-95, improving by 2.8% and 3.3% over YOLO11n baseline respectively. With 6.6 GFLOPs and 2.51M parameters, HMIV-DET achieves 93.0% precision and 87.8% recall, exceeding YOLOv11n by 0.9% and 3.6% respectively while maintaining lightweight characteristics. Compared to specialized small object detection methods, HMIV-DET demonstrates substantial advantages. Against TPH-YOLOv5, HMIV-DET achieves 2.6% higher mAP50 while requiring 36 times fewer computational resources and 35 times smaller model size. HMIV-DET outperforms UAV-DETR by 3.4% in mAP50 and SO-DETR by 2.7% in mAP50, with computational costs reduced by 13 times and 11 times respectively. Most critically, HMIV-DET achieves 58.5% APS, outperforming TPH-YOLOv5 by 2.7%, UAV-DETR by 3.0%, and SO-DETR by 1.7%, validating that AIVE’s mirror-aware preprocessing addresses spatial unfairness inherent in indirect vision systems. DETR-based methods employ transformer architectures and query-based detection for small objects but lack mechanisms to handle distance-dependent resolution degradation across multiple mirror regions. Their substantial computational overhead limits real-time deployment. HMIV-DET achieves balanced multi-scale performance with 66.2% APM, surpassing all comparison methods. This cross-scale robustness confirms the synergistic effectiveness of our three innovations: AIVE establishes spatial fairness, AMFOBlock captures diverse flame morphologies, and HCSFPN maintains detection capability across all scales through attention-driven fusion. Visualization analysis Figure 8 employs Grad-CAM visualization to reveal feature saliency distribution differences among YOLOv8, YOLO11, and HMIV-DET in indirect-view fire detection. Grad-CAM feature activation heatmap visualization comparing YOLOv8, YOLO11, and HMIV-DET across four mirror regions. Scenario (a) upper-left mirror shows HMIV-DET’s compact high-intensity activation precisely covering flame core, while baseline methods exhibit diffuse activation patterns extending into backgrounds. Scenarios (b-c) upper-right and lower-right mirrors demonstrate HMIV-DET’s consistent activation concentration across spatial locations with heat zones confined to actual flame boundaries, validating HCSFPN’s position-invariant discrimination through attention-driven fusion. Scenario (d) lower-left mirror confirms superior spatial compactness and peak response intensity. The progressive evolution from YOLOv8’s coarse-grained global activation to YOLO11’s moderate focusing to HMIV-DET’s fine-grained target-specific activation validates the synergistic effect of AIVE preprocessing, AMFOBlock’s feature extraction, and HCSFPN’s semantic fusion in constructing an enhanced feature representation space. The activation response for the upper-left mirror flame target in scenario (a) shows that HMIV-DET exhibits a compact high-intensity activation region (heat zone precisely covering the flame core), whereas YOLOv8 and YOLO11 show diffuse activation patterns with boundaries extending into surrounding backgrounds, validating that the feature-level suppression mechanism implemented by AMFOBlock through gated selective activation units enhances target-background discrimination capability. Scenarios (b) and (c) demonstrate flame detection in the upper-right and lower-right mirrors respectively. HMIV-DET maintains consistent activation concentration across different spatial locations, with heat zones confined to actual flame boundaries and background regions displaying deeper blue coloration, while baseline methods show activation spillover and redundant weak activations in non-target mirror areas. This validates from the feature perspective that cross-scale semantic propagation achieved by HCSFPN through attention-driven fusion equips the network with position-invariant discrimination capability. The lower-left mirror flame in scenario (d) further confirms this advantage, as HMIV-DET’s activation heat zone exhibits higher spatial compactness and stronger peak response intensity. The evolution from YOLOv8’s coarse-grained global activation to YOLO11’s moderate focusing and then to HMIV-DET’s fine-grained target-specific activation, confirming from a neural network interpretability perspective that AIVE’s view equity preprocessing, AMFOBlock’s multi-geometric adaptive feature extraction, and HCSFPN’s hierarchical semantic fusion synergistically construct a feature representation space with enhanced discrimination and robustness. Conclusions This paper addresses special challenges brought by indirect vision technology in heritage building fire monitoring through the HMIV-DET algorithm. The AIVE module resolves spatial unfairness in multi-mirror deployments by ensuring remote small-size mirror regions receive sufficient attention through dynamic weight allocation. AMFOBlock overcomes traditional fixed convolution kernel limitations by integrating multi-geometric adaptive convolution and gated selective activation for complex flame morphology. HCSFPN addresses limited semantic information propagation through dense connections and attention-driven multi-scale fusion strategies. Experimental results validate the method’s effectiveness, with mAP50 and mAP50-95 improving by 2.8% and 3.3% respectively compared to YOLO11 baseline on the Heritage-IVF dataset. Comprehensive experiments demonstrate AIVE’s improvement of 1.4% in mAP50 over the best competitor, AMFOBlock’s highest performance with 6.3% GFLOPs reduction, and HCSFPN’s effectiveness over alternative architectures. However, this study has certain limitations. Experiments primarily use virtual fire scenes constructed with 3DMAX software, and while 3D modeling strictly matches actual heritage structures and environmental parameters, simulated fires may differ from real scenarios in combustion dynamics and smoke behavior. The algorithm’s performance can be affected by mirror surface degradation, as enhancement processing cannot recover information lost to physical obstruction. Current methods process individual frames independently without leveraging temporal information. Performance may degrade under extreme conditions due to fundamental physical and resolution constraints. Deployment on resource-constrained embedded systems remains challenging despite efficiency improvements. Future work will address these challenges through constructing large-scale real-world fire datasets via controlled experiments and historical footage collaboration, developing adaptive mirror quality assessment and temporal processing mechanisms, integrating multi-modal sensor fusion for improved robustness, and advancing model compression techniques for practical embedded deployment. This work provides important technical support for heritage building intelligent fire monitoring systems, while the acknowledged limitations and proposed future directions establish a foundation for advancing toward increasingly robust heritage fire protection solutions. Author contributions J.L. proposed the HMIV-DET algorithm’s core modules (AIVE, AMFOBlock, HCSFPN), constructed the Heritage-IVF dataset, conducted experiments, and drafted the manuscript. J.C. (corresponding author) supervised the research design, reviewed the experimental scheme, and revised the manuscript. All authors reviewed and approved the final version of the manuscript. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Data availability The datasets generated during and/or analysed during the current study are not publicly available but are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets generated during and/or analysed during the current study are not publicly available but are available from the corresponding author on reasonable request.