Enhancing Wireless Networks for IoT with Large Vision Models: Foundations and Applications

Yunting Xu, Jiacheng Wang, Ruichen Zhang, Dusit Niyato, Deepu Rajan, Liang Yu · 2025

Read Enhancing Wireless Networks for IoT with Large Vision Models: Foundations and Applications by Yunting Xu, Jiacheng Wang, Ruichen Zhang, Dusit Niyato, Deepu Rajan, Liang Yu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-00583.txt.

Introduction

Yunting Xu, Jiacheng Wang, and colleagues survey how large vision models can strengthen wireless networks for connected devices in a 2025 study. Vision foundation models now generalize across many visual tasks, and that adaptability suits network jobs such as reading a scene to aim beams or predict blockages. The authors first lay out what these models are and how their architectures work.

Applications across the Internet of Things follow, from camera-assisted configuration to vision-guided optimization. The survey weighs where a big model's flexibility beats a small specialist. It gives network engineers a map for borrowing progress from computer vision.

Catalog record

CREATORYunting Xu, Jiacheng Wang, Ruichen Zhang, Dusit Niyato, Deepu Rajan, Liang Yu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-00583

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Enhancing Wireless Networks for IoT with Large Vision Models: Foundations and Applications Thanks: Y. Xu, J. Wang, R. Zhang, D. Niyato, and D. Rajan are with the College of Computing and Data Science, Nanyang Technological University, Singapore, 639798 (e-mail: [email protected], [email protected], [email protected], [email protected], [email protected]). Thanks: L. Yu is with Alibaba Cloud, China (e-mail: [email protected]). Thanks: H. Zhou is with the School of Electronic Science and Engineering, Nanjing University, Nanjing, China, 210023 (e-mail: [email protected]). Thanks: A. Jamalipour is with The University of Sydney, Sydney NSW 2006, Australia (e-mail: [email protected]). Thanks: X. Wang is with the Department of Electrical and Computer Engineering, Western University, Canada (e-mail: [email protected]). Yunting Xu, Jiacheng Wang, Ruichen Zhang, Dusit Niyato, Deepu Rajan, Liang Yu Abstract Large vision models (LVMs) have emerged as a foundational paradigm in visual intelligence, achieving state-of-the-art performance across diverse visual tasks. Recent advances in LVMs have facilitated their integration into Internet of Things (IoT) scenarios, offering superior generalization and adaptability for vision-assisted network optimization. In this paper, we first investigate the functionalities and core architectures of LVMs, highlighting their capabilities across classification, segmentation, generation, and multimodal visual processing. We then explore a variety of LVM applications in wireless communications, covering representative tasks across the physical layer, network layer, and application layer. Furthermore, given the substantial model size of LVMs and the challenges of model retraining in wireless domains, we propose a progressive fine-tuning framework that incrementally adapts pretrained LVMs for joint optimization of multiple IoT tasks. A case study in low-altitude economy networks (LAENets) demonstrates the effectiveness of the proposed framework over conventional CNNs in joint beamforming and positioning tasks for Internet of drones, underscoring a promising direction for integrating LVMs into intelligent wireless systems. I Introduction In recent years, the integration of foundation models into visual tasks for Internet of Things (IoT) scenarios has emerged as a prevailing paradigm, exemplified by the deployment of large vision models (LVMs) such as the Vision Transformer (ViT) [ 1 ] . Different from large language models (LLMs) that primarily focus on sequential textual data, LVMs are specifically tailored for spatial and structural comprehension of visual inputs, enabling state-of-the-art performance across IoT tasks such as image recognition, semantic segmentation, and vision-language reasoning. Owing to the massive number of parameters and extensive pretraining on large-scale visual datasets, LVMs consistently outperform traditional convolutional neural networks (CNNs) in representation learning and model generalization. Furthermore, their substantial scalability and cross-domain adaptability facilitate effective knowledge transfer across a wide range of downstream applications [ 2 ] . Among these emerging application domains, LVM-based wireless communications have attracted considerable attention, offering promising opportunities to improve communication performance in complex and dynamic environments [ 3 ] . Leveraging their advanced visual understanding capabilities, LVMs extract rich contextual information from visual inputs, enabling a range of vision-assisted communication tasks such as channel estimation, beamforming, near-field signal processing, and semantic-aware resource allocation [ 4 ] . The vision-derived information offers a complementary alternative to conventional radio frequency measurements, thereby significantly enhancing the system’s perception of complex wireless environments. Moreover, LVMs demonstrate robust generalization and adaptability for the simultaneous optimization of multiple IoT tasks. The capability underscores the potential of LVMs to serve as a unified foundational backbone for vision-assisted, multi-task wireless communication systems. Despite the promising performance of LVMs in IoT task optimization, several implementation challenges still remain in realizing their full potential for wireless deployment scenarios. On the one hand, the large model size of LVMs leads to significant computational overhead and energy consumption during the training process, making full retraining from scratch both resource-intensive and challenging to converge. On the other hand, IoT tasks in wireless communication environments often suffer from a lack of task-specific and large-scale labeled datasets, which limits the ability to effectively adapt LVMs to domain-specific scenarios. These constraints hinder the direct deployment of LVMs in wireless networks and necessitate efficient adaptation strategies tailored to data-scarce and resource-constrained environments. Motivated by the increasing imperative to exploit the full potential of LVMs in wireless IoT task optimization, this paper contributes to the field through investigating the architectures and functionalities of LVMs. To address the practical challenges of adapting LVMs to wireless networks, we redesign the model’s output structure to facilitate joint optimization across multiple wireless communication tasks. Furthermore, we propose a progressive fine-tuning framework that incrementally unfreezes the LVM layers during the training process. The main contributions of this paper are summarized as follows. • We investigate state-of-the-art LVM architectures and analyze their functional characteristics in the context of visual understanding tasks, highlighting their generalization and adaptability across diverse vision applications. • We systematically explore the integration of LVMs into wireless networks, with emphasis on their utility across the physical layer, network layer, vision-based semantic communications, and user-centric service applications. • We propose a fine-tuning framework for LVMs based on a multi-stage unfreezing pipeline, designed to support wireless multi-task optimization. A case study conducted in low-altitude economy networks (LAENets) demonstrates the performance advantages of the proposed framework over traditional CNN approaches for Internet of drones. II Foundations of large vision model This section examines the architectural and functional properties of LVMs. Representative designs and application domains are visualized in Figure 1 , and a detailed overview of state-of-the-art LVMs is provided in Table I . II-A Functionality of Large Vision Models II-A 1 Image Classification Image classification aims to predict the category of input images. It directly processes the entire image and outputs a single class label, identifying the primary object or scene type. Unlike conventional CNN-based approaches, LVMs leverage transformer-based architectures, such as ViT and Swin Transformer, which capture global spatial patterns through patch-based self-attention mechanisms. These models offer superior accuracy and robustness when handling diverse and complex image datasets. II-A 2 Object Detection Object detection involves identifying and localizing multiple discrete entities within an image. Detection models output precise spatial coordinates, represented as bounding boxes, along with semantic labels and associated confidence scores. Recent models such as DETR leverage an encoder-decoder-based transformer architecture to enable efficient multi-object detection. This functionality enhances both spatial and semantic understanding, facilitating applications such as autonomous navigation and real-time monitoring. II-A 3 Image Segmentation Image segmentation partitions an image into coherent, semantically meaningful regions at a pixel-level resolution. Recent LVMs, exemplified by Segment Anything Model (SAM), employ prompt-based, category-agnostic methods to effectively segment arbitrary and unseen objects. The generalization capability of LVM significantly extends segmentation versatility and applicability across diverse and previously unencountered scenarios. II-A 4 Image reconstruction Image reconstruction tasks focus on recovering missing, corrupted, or partially occluded image content. Leveraging self-supervised learning approaches, advanced LVMs such as Masked Autoencoder (MAE) employ Autoencoder to reconstruct occluded image regions through sophisticated masked encoding and decoding frameworks. The self-supervised pretraining paradigm of LVM enhances representation robustness, enabling effective learning from limited or incomplete data, even in the absence of labels. II-A 5 Image representation Visual representation learning seeks to extract semantically rich, transferable feature embeddings from images. LVMs such as DINO employ self-distillation techniques to efficiently generate meaningful visual representations without the need for explicit annotations. The representation capability of LVM can significantly enhance downstream performance across various vision-assisted tasks. II-A 6 Generation Capability Generative LVMs synthesize novel content based on learned patterns and data

Read the complete text