CANNE: CLIP-Based ANNE Selection for Noisy-Label Learning Authors: Ge Jin, Qian Zhang, Li Huang, Yongqi Yuan, Yulong Qiao, De Yu, Mengjie Ye, Yingwen Zhu Published: 2026 ABSTRACT Learning with noisy labels (LNL) remains challenging, especially when the identification of clean samples relies heavily on the predictions of the model being trained. In such cases, early-stage selection errors may be reinforced during iterative optimization, leading to unreliable supervision. To alleviate this issue, a two-stage framework, termed CANNE, is proposed by combining Contrastive Language–Image Pre-training (CLIP)-based conservative offline cleaning with Adaptive Nearest Neighbors and eigenvector-based sample selection (ANNE)-based online refinement. Specifically, a high-confidence clean seed set is first constructed using two complementary probability sources derived from frozen CLIP representations and reliability criteria, including class-wise loss modeling and prediction consistency. This seed set is then used as a set of reliable anchors during the subsequent ANNE training process, where online feature- and neighborhood-based refinement further recovers and adjusts sample partitions. In this way, CANNE uses external vision–language priors to provide conservative and persistent guidance while preserving the adaptive recovery ability of online noisy-label learning. Experimental results on CIFAR-10, CIFAR-100, Animal-10N, and Mini-WebVision, together with additional evaluation under open-set noise, show that the proposed method achieves competitive performance across diverse noisy-label settings. In particular, CANNE achieves 96.6% and 96.3% best accuracies on CIFAR-10 under 80% and 90% symmetric noise, respectively, and 81.0% and 79.0% on CIFAR-100 under 20% and 50% symmetric noise. Additional repeated-run, threshold-sensitivity, and runtime analyses further indicate that the CLIP-based seed set provides stable guidance with only moderate computational overhead. 1. Introduction The rapid advancement of deep neural networks (DNNs) in computer vision has been largely supported by the availability of large-scale annotated datasets. However, collecting high-quality labels at scale is often costly and time-consuming. In practice, many datasets are therefore collected from web queries or crowdsourcing platforms, where label noise is almost unavoidable. Training directly on such corrupted annotations often leads to noticeable performance degradation, as over-parameterized DNNs tend to gradually memorize noisy labels during optimization [1]. Consequently, learning with noisy labels (LNL) has become an important problem in building robust visual recognition systems. Over the past few years, a variety of strategies have been developed to address label noise, among which sample selection has emerged as one of the most effective directions. The main idea is to identify a subset of likely clean samples during training, often based on the small-loss phenomenon that DNNs tend to fit clean patterns earlier than noisy ones [2]. Despite substantial progress, most existing selection methods still rely heavily on the predictions of the network being trained. This gives rise to a practical chicken-and-egg dilemma: robust classification requires reliable clean samples, while the identification of clean samples itself depends on the reliability of the classifier. During the early stages of training, the model may remain insufficiently discriminative, causing false sample selection. Once such errors are fed back into the optimization loop, confirmation bias can gradually accumulate and eventually impair performance, especially under severe or open-set noise conditions. To mitigate this issue, additional information that is less coupled with the noisy labels of the target training set can be beneficial. Recently, vision–language models (VLMs), such as CLIP [3], have shown strong zero-shot recognition ability by learning transferable image–text representations from large-scale pre-training. Since CLIP is trained independently of the target noisy dataset, its semantic predictions are not directly produced by the in-training classifier and can therefore provide an external source of reliability estimation. This property has motivated recent studies, such as CLIPCleaner [4], to use CLIP for offline clean-sample selection in noisy-label learning. CLIPCleaner demonstrates that CLIP can provide a useful external reliability estimate for identifying likely clean samples. Nevertheless, CLIP-based cleaning alone is not always sufficient for the entire noisy-label learning process. Zero-shot predictions may be affected by domain gaps, fine-grained category ambiguity, or prompt sensitivity. Moreover, conservative CLIP-based selection usually prioritizes precision and may leave many clean but difficult samples unselected. These observations suggest that CLIP-based offline cleaning and online noisy-label refinement are complementary: the former provides reliable external seeds, while the latter can adapt to the target dataset and recover additional samples during training. Motivated by this complementarity, a two-stage framework termed CANNE is proposed to connect conservative CLIP-based offline cleaning with ANNE-based online refinement. In the first stage, a high-confidence clean seed set is constructed using CLIP-derived reliability estimates. This seed set is intentionally conservative: it is expected to contain highly reliable samples, while some clean samples may remain unselected. In the second stage, these seeds are incorporated into ANNE [5], which dynamically refines sample partitions according to feature structure and neighborhood relationships. The CLIP-selected seeds act as reliable anchors during online refinement, helping reduce early-stage error accumulation while allowing the training model to adaptively recover additional samples from the remaining data. The main contributions of this work are summarized as follows. A two-stage noisy-label learning framework, termed CANNE, is proposed to integrate CLIP-based conservative offline clean-seed construction with ANNE-based online adaptive refinement. The offline stage combines two complementary reliability estimates derived from frozen CLIP representations, namely zero-shot image–text probabilities and visual-feature-induced probabilities, thereby reducing the exclusive dependence of early sample selection on the in-training classifier. A persistent seed-guided refinement strategy is developed to preserve high-confidence CLIP-selected samples as reliable anchors during ANNE training, while samples outside the seed set remain subject to adaptive online selection. This design provides stable supervision during early refinement without removing the feature- and neighborhood-based sample recovery ability of ANNE. Extensive experiments on CIFAR-10, CIFAR-100, Animal-10N, and Mini-WebVision, together with additional evaluation under open-set noise, show that the proposed method achieves competitive performance across diverse noisy-label settings, with particularly favorable results in several severe-noise scenarios. 2. Related Work 2.1. Learning with Noisy Labels Learning with noisy labels has been widely studied because large-scale training data collected from web search engines, crowdsourcing platforms, or weakly supervised pipelines usually contain imperfect annotations. A central difficulty in this problem is that deep neural networks can gradually memorize corrupted labels during optimization [1]. As a result, standard training with cross-entropy loss often leads to degraded generalization when the noise ratio becomes high or the noise pattern becomes complicated. Several types of methods have been developed to reduce the effect of noisy supervision. Robust loss functions attempt to make the training objective less sensitive to corrupted labels. Early studies analyzed learning under noisy labels from a risk correction perspective [6,7]. Generalized cross entropy was introduced to balance the optimization advantage of cross entropy and the robustness of mean absolute error [8], while symmetric cross entropy used a reverse cross-entropy term to improve noise tolerance [9]. Other objectives, such as normalized losses and active negative losses, were also investigated for more robust optimization under label noise [10,11]. Prediction-consistency regularization combined with contrastive clustering has also been explored to improve robustness to noisy supervision [12]. MixUp has been commonly used to regularize the decision boundary and reduce overfitting to individual noisy samples [13], and its relation to the vicinal risk minimization principle has also been studied theoretically [14]. These methods are usually simple and can be combined with other frameworks, but their standalone performance may be limited under severe noise. Label correction is another important direction. Instead of directly discarding unreliable samples, noisy annotations are corrected or refined according to model predictions or estimated noise patterns. Joint optimization methods update labels and network parameters simultaneously [15], while probabilistic end-to-end correction methods refine noisy labels during training [16]. Progressive correction has also been studied for feature-dependent label noise [17]. More recent methods further combine label correction with contrastive or prototypical representations to improve the quality of corrected labels [18,19,20]. Other approaches, including reflective learning [21], part-level noisy-class posterior estimation [22], and uncertainty-guided label correction [23], provide additional mechanisms for refining noisy supervision. Although label correction can recover part of the corrupted supervision, its reliability is still closely tied to the quality of the model predictions. When the classifier is not sufficiently reliable, especially in early training, incorrect corrections may be introduced and reinforced. Sample selection has therefore become one of the most influential strategies in LNL. The basic idea is to identify likely clean samples and use them as reliable supervision. Many sample-selection methods are based on the small-loss assumption, which indicates that clean samples are typically learned earlier and thus have lower losses than noisy samples at the beginning of training [24]. Co-teaching follows this principle by training two networks and allowing them to exchange selected small-loss samples [25]. DivideMix extends this strategy by fitting the sample-wise loss distribution with a Gaussian mixture model and treating samples with unreliable labels as unlabeled data in a semi-supervised learning framework [2]. Later studies improved sample selection through high-confidence sample mining [26], cyclical joint-loss and co-teaching strategies [27], dynamic instance-specific selection and correction [28], optimal transport-based filtering [29], uniform sample selection with contrastive learning [30], hyperspherical margin weighting [31], learning-risk-based separation of hard clean and noisy samples [32], and agreement-based co-regularization [33]. These methods have shown strong performance in many noisy-label settings. However, most sample-selection methods still estimate sample reliability from the network being optimized, using its losses, predictions, or learned features. This dependency can be problematic in the early training stage, when the model is not yet discriminative. Once incorrectly selected samples are fed back into the training loop, the resulting confirmation bias may affect subsequent optimization. This concern is related to recent studies on prediction instability. Jauhari and Hidey [34] studied compute-efficient churn reduction for conversational agents by reducing undesirable output changes between model versions, while Datta et al. [35] measured and mitigated local instability in deep neural networks under input perturbations. Although their settings differ from LNL, these studies highlight that unstable model decisions can affect downstream decision-making. In noisy-label learning, such instability is particularly harmful because early incorrect clean/noisy partitions may be reinforced during training. This limitation motivates the use of additional information that is less coupled with the noisy labels of the target training set. 2.2. Feature-Based Sample Selection and Online Refinement In addition to loss-based sample selection, feature-based selection has also been explored. Instead of relying only on classification losses or prediction confidence, these methods use the structure of the learned feature space to distinguish clean and noisy samples. The underlying assumption is that samples from the same semantic class tend to form compact or structured clusters, whereas mislabeled samples are more likely to deviate from the dominant class structure. FINE detects clean samples by measuring the alignment between sample representations and class-wise dominant eigen-directions [36]. The dominant eigenvector of the feature distribution is used as a class-specific reference direction, and samples that are more consistent with this direction are more likely to be selected as clean. This provides a feature-space criterion beyond the loss value. Nevertheless, the quality of the selection still depends on whether the learned representation has become sufficiently meaningful. Neighborhood-based methods further exploit local relationships in the feature space. SSR introduces a sample selection and relabeling framework based on nearest-neighbor voting and confidence-based relabeling [37]. In this framework, the local neighborhood structure provides evidence for whether a sample label is reliable. ANNE further follows a heterogeneous selection design by combining different strategies for samples with different confidence levels [5]. The training set is first partitioned into high-confidence and low-confidence subsets. Eigenvector-based selection is then applied to the high-confidence subset, while adaptive nearest-neighbor selection is used for the low-confidence subset. This heterogeneous design is useful because samples with different confidence levels may require different selection rules. Several recent methods also adopt progressive or two-stage designs. LongReMix constructs a high-confidence clean set and uses it to guide subsequent robust training [26]. C2MT introduces a cross-to-merge training strategy with a class balance mechanism to improve robustness under noisy supervision [38]. PSSCL decouples clean-set discovery and robust training through progressive sample selection with contrastive learning [39]. These methods suggest that reliable initialization and adaptive refinement are both important for noisy-label learning. TCC-net also adopts a two-stage design by combining contradictory loss, co-teaching, and meta-learning for training under noisy supervision [40]. More broadly, contrastive representation learning has provided an important foundation for robust feature learning. SimCLR [41] and MoCo [42] established representative self-supervised contrastive frameworks, while supervised contrastive learning extended this principle to label-aware representation learning [43]. These ideas were subsequently adapted to webly or noisily supervised learning through momentum prototypes [44], selective supervised contrastive learning [45], and semantic clustering with semi-supervised learning [46]. Despite these advances, the initial partitioning and refinement process in these methods is still mostly driven by the in-training model. When early predictions or representations are unreliable, the online selection process may start from a biased state. This observation motivates the use of an external clean seed set to guide the online refinement stage. In this work, ANNE is adopted as the online refinement framework because it already provides a heterogeneous selection mechanism based on feature structure and neighborhood relationships. The main difference is that an offline CLIP-based clean seed set is injected before online refinement so that ANNE can be initialized with more reliable supervision. 2.3. Vision–Language Priors for Noisy-Label Cleaning Recently, vision–language models have provided a new source of external semantic knowledge. CLIP learns transferable visual representations from large-scale image–text pairs and has shown strong zero-shot recognition ability [3]. Since CLIP is pre-trained independently of the target noisy dataset, its predictions are less directly affected by the corrupted labels used during downstream training. This property makes CLIP attractive for noisy-label cleaning, where the dependence on the in-training classifier is one of the main sources of confirmation bias. CLIPCleaner is one of the first works to use CLIP for noisy-label sample selection [4]. A CLIP-based zero-shot classifier is constructed, and clean samples are selected according to complementary criteria, including prediction consistency and loss modeling based on CLIP-induced probabilities. Since the selection process is performed offline and is decoupled from the downstream classifier, the influence of noisy labels on the sample selection stage can be reduced. This provides a useful external prior for LNL. However, CLIP-based zero-shot predictions are not always sufficient for the whole noisy-label learning process. Domain gaps may exist between the CLIP pre-training data and the target dataset. Fine-grained categories or visually ambiguous classes may also reduce the reliability of zero-shot predictions. In addition, a conservative CLIP-based selector usually prioritizes precision and may have limited recall, which means that many clean but difficult samples may not be selected at the initialization stage. These observations indicate that CLIP-based cleaning is more suitable as a reliable initialization mechanism than as a complete replacement for adaptive noisy-label learning. A CLIP-based selector can provide a high-precision subset that is less affected by the in-training classifier, but its conservative nature may limit sample coverage. Online refinement methods, in contrast, can adapt to the target data distribution and recover additional samples, but their early decisions may be affected by unreliable model predictions. CANNE is built on this complementarity. The CLIP-based stage supplies a high-confidence seed set, and the ANNE stage further refines sample partitions using the evolving feature and neighborhood structure of the target model. 3. Method 3.1. Formulation Consider a noisy training dataset D={(xi,y~i)}i=1N with c classes, where xi∈R3×H×W denotes an input image and y~i∈{1,…,C} denotes its observed label, which may be corrupted. Here, H and W indicate the height and width of the image, respectively. The corresponding clean ground-truth label is denoted by yi∈{0,1}c. Due to the presence of annotation noise during the data collection, cases where yi≠y~i are common in the dataset D. The goal of LNL is to train a robust DNN under such noisy conditions. Assume that the DNN consists of a feature extractor fϕ and a classifier gθ, where φ and θ are the parameters, respectively. For an input sample xi, its feature representation is given by zi=fϕxi, and the prediction on the probability simplex is pi=gθzi, where pi∈0, 1c The corresponding hard label prediction is denoted as y^i=arg maxk pi(k). The proposed method, CANNE, is designed to address three common types of label noise: symmetric noise, asymmetric noise, and open-set noise. Symmetric noise is generated by replacing the clean label with a label sampled uniformly from the label space according to a given noise ratio, following the commonly used noisy-label protocol. Asymmetric noise, on the other hand, denotes the case where a clean label is flipped, with probability η, only to visually similar classes (i.e., similar-looking classes). This type of noise is class-dependent, meaning that the transition probabilities depend on the class rather than the input features. Open-set noise is one of the most challenging types, as the true class of a sample may not belong to the predefined label space of the dataset, i.e., the sample may originate from unknown classes outside the training distribution. The overall framework of CANNE is illustrated in Figure 1. The proposed method consists of two main stages. In the first stage, CLIP is used as an external vision–language prior to perform conservative offline sample cleaning. Specifically, class-wise loss modeling and prediction consistency filtering are jointly used to construct a high-confidence clean seed set D_clip. In the second stage, this clean seed set is injected into the ANNE-based online refinement process. ANNE dynamically partitions the training samples into high-confidence and low-confidence subsets and applies heterogeneous selection strategies to identify clean and noisy samples. By combining the offline CLIP-based clean seeds with online feature-based refinement, CANNE aims to reduce early-stage confirmation bias while preserving the adaptive sample recovery ability of ANNE. 3.2. CLIP-Based Offline Sample Cleaning To reduce the dependence of early sample selection on the in-training model, a CLIP-based offline sample cleaning module, denoted as CSC, is applied before ANNE training. The design follows the conservative sample-selection principle of CLIPCleaner [4], where CLIP is used as an external vision–language model to estimate the compatibility between each image and its observed label. In CANNE, the selected samples are not treated as the final clean training set. Instead, they form a high-confidence seed set that provides reliable anchors for the subsequent ANNE-based online refinement. Let D=(xi,y~i)i=1N denote the noisy training set, where xi is the i-th image and y~i∈{1,...,C} is its observed label. Here, C is the number of classes. CLIP consists of an image encoder Eθv and a text encoder Eθt. In our implementation, the OpenAI CLIP model with the ViT-B/32 visual backbone is used by default. CLIP is used only for inference, and all its parameters are frozen throughout the whole process. No CLIP fine-tuning, prompt tuning, or knowledge distillation is performed. Images are processed by the standard preprocessing pipeline returned by the OpenAI CLIP implementation, including resizing, center cropping, and CLIP normalization. This keeps the offline cleaning stage independent of the noisy-label training process. For each class c, a set of descriptive prompts {Pc,j}j=1J is constructed. The prompt template follows the form “a photo of a {class name} {class-specific description}”. The class-specific descriptions are short semantic attributes associated with the class, such as object parts, shapes, or typical visual properties. For example, prompts may include “a photo of an airplane which has wings and a fuselage”, “a photo of an automobile which has four wheels”, “a photo of a ship which is a vessel on water”, and “a photo of a cat which has whiskers and pointed ears”. The prompt set is fixed before training and is not optimized using the noisy labels. For an image xi and the j-th prompt Pc,j of class c, CLIP computes the image–text similarity through the inner product between the encoded image and text features. Following the standard CLIP inference procedure, both image and text features are L2-normalized before computing the similarity. For notational simplicity, Eθv(xi) and Eθt(Pc,j) are used below to denote the normalized image and text features, respectively. The multi-prompt class score for class c is defined as:(1)si,czs=∑j=1JexpEθv(xi)⊤Eθt(Pc,j)T, where T is the CLIP temperature and is set to 0.07. The superscript “zs” denotes the zero-shot probability source. The zero-shot conditional probability is then obtained by normalizing the class scores over all C classes:(2)pi,czs=P~zs(y=c|xi)=∑j=1JexpEθv(xi)⊤Eθt(Pc,j)/T∑k=1C∑j=1JexpEθv(xi)⊤Eθt(Pk,j)/T. Here, Pc,j represents the j-th prompt of class c, and J is the number of prompts per class. The probability vector of sample x_i obtained from CLIP zero-shot prediction is denoted as pizs=[pi,1zs,...,pi,Czs]. In addition to the zero-shot probability pizs, CSC also uses an auxiliary probability estimated from frozen CLIP visual features. Specifically, a balanced logistic regression classifier is trained on the frozen CLIP image features using the observed labels, and its output probability vector is denoted as pilr. This induced classifier is used only in the offline cleaning stage and does not update CLIP. The zero-shot probability pizs and the induced probability pilr provide two complementary reliability estimates: the former exploits image–text semantic alignment, while the latter adapts frozen CLIP visual representations to the target label space. The two probability sources are not assumed to be individually reliable enough to define the final seed set. The zero-shot source is independent of the observed training-label distribution but can be affected by prompt wording and image–text domain mismatch. The visual-LR source adapts the frozen visual representation to the target classes but is trained using the observed noisy labels and can therefore inherit label corruption. CSC uses their intersection because a sample supported by both sources is less likely to be selected due to a source-specific error. The contribution of each source is examined through controlled experiments in Section 4.4.3. Based on the CLIP-derived probability pir, where r∈{zs,lr}, two complementary sample selection criteria are introduced: class-wise loss modeling and prediction consistency filtering. These two criteria evaluate label reliability from different perspectives. The loss-based criterion measures the compatibility between the observed label and the estimated class probability, while the consistency criterion checks whether the observed label agrees with the most confident prediction. The final CSC seed set is obtained by conservatively intersecting the selected subsets so that precision is prioritized at the initialization stage. (a)Loss-based filtering with class-wise Gaussian mixture modeling Loss-based filtering with class-wise Gaussian mixture modeling The first criterion evaluates how compatible the observed label is with the CLIP-derived prediction. For each CLIP-derived probability vector pir, where r∈zs,lr, the cross-entropy loss score of sample xi with respect to its observed label y~i is defined as(3)lir=−logpi,y~ir. Clean samples are expected to have smaller loss values because their observed labels are more consistent with the CLIP-derived prediction. To reduce the influence of class imbalance and inter-class score-scale differences, the loss distribution is modeled separately for each observed class. For class c, let Ic=i|y~i=c denote the index set of samples whose observed label is c. The loss values lir|i∈Ic are assumed to follow a bimodal distribution corresponding to clean and noisy samples and are fitted by a two-component Gaussian mixture model:(4)p(lir|y~i=c)=∑m=12πc,mrNlir;μc,mr,(σc,mr)2, where πc,mr,μc,mr, and σc,mr denote the mixture weight, mean, and standard deviation of the m-th component for class c, respectively. When the score is written as the cross-entropy loss, the component with the smaller mean is regarded as the clean component. Let mcr=argminm∈{1,2}μc,mr denote the clean component for class c. The posterior probability that sample xi belongs to this clean component is computed as(5)qir,gmm=πc,mcrrNlir;μc,mcrr,(σc,mcrr)2∑m=12πc,mrNlir;μc,mr,(σc,mr)2, where c is the observed class satisfying y~i=c. A sample is selected by the loss-based criterion if(6)qir,gmm≥τgmm, where τgmm is the GMM posterior threshold. Unless otherwise specified, τgmm is set to 0.5 in the main experiments. In implementation, the class-wise scores are normalized before GMM fitting to improve numerical stability. Equivalently, when the normalized log-probability confidence score is used instead of the cross-entropy loss, the clean component corresponds to the component with the larger mean, which is monotonic to the negative cross-entropy loss. This does not change the selection principle. Since the evaluated datasets contain sufficient observed samples for each class, the class-wise GMM fitting is stable in our experiments. In the reported experiments, the conservative intersection produces non-empty seed sets for all classes. Therefore, no additional test-set information or manual class-specific adjustment is used. (b)Prediction consistency filtering with a fixed threshold Prediction consistency filtering with a fixed threshold The second criterion evaluates the consistency between the observed label and the most confident CLIP-derived prediction. For each probability source pir, where r∈{zs,lr}, the predicted class of sample xi is defined as(7)y^ir=argmaxk∈{1,…,C}pi,kr. The consistency score is then computed as(8)air=pi,y~irmaxk∈{1,…,C}pi,kr. This score measures the relative agreement between the observed label and the most confident prediction. If the observed label is consistent with the dominant prediction, air tends to be close to 1; otherwise, it becomes smaller. A sample is selected as clean by the consistency criterion if(9)air≥τcons, where τcons is a fixed consistency threshold. Unless otherwise specified, τcons is set to 0.8 in the main experiments. Finally, to improve seed precision, CSC conservatively aggregates the selected samples by intersection. Specifically, for both the CLIP zero-shot probability and the CLIP-feature-induced probability, a sample is retained only when it satisfies both the loss-based GMM criterion and the prediction consistency criterion. The resulting high-confidence clean subset is denoted as Dclip, which is used as the initial clean seed set for the subsequent ANNE-based online refinement. This conservative strategy prioritizes precision over recall, while clean but difficult samples not selected at this stage can still be recovered during the online refinement stage. 3.3. Integration with ANNE Framework After obtaining the initial clean seed set Dclip from CSC, we integrate it into the ANNE [5] framework to achieve a fusion of offline and online sample selection. The CLIP-selected subset is not used to replace the original noisy training set. Instead, it provides high-confidence seed anchors for the subsequent online refinement. During the ANNE stage, the entire noisy training set is still processed, and additional reliable samples are progressively identified by jointly exploiting model confidence and feature structure. Specifically, let fϕ denote the feature extractor and gθ denote the classifier. For each sample xi, the network predicts a softmax probability vector pi=gθ(fϕ(xi)). Based on the maximum posterior confidence, the noisy training set is partitioned into a high-confidence set D_HCS and a low-confidence set D_LCS as(10)DHCS=(xi,y~i)∈D | maxk∈{1,…,C}pi,k≥τconfDLCS=(xi,y~i)∈D | maxk∈{1,…,C}pi,k<τconf, Here, τconf is estimated by Otsu’s thresholding criterion, which determines the threshold by maximizing the between-group variance of the confidence scores. After the confidence-based partition, two complementary online selection strategies are applied to DHCS and DLCS, respectively. (a)High-Confidence Set (HCS) High-Confidence Set (HCS) Samples in DHCS are expected to contain a relatively high proportion of clean labels. Therefore, a feature decomposition strategy based on class-wise principal directions is adopted. For each class c, let IcHCS=i | (xi,y~i)∈DHCS, y~i=c denote the index set of high-confidence samples whose observed label is c. The normalized feature of sample xi is defined as zi=fϕ(xi)/||fϕ(xi)||2. A class-wise scatter matrix is constructed as(11)Cc=∑i∈IcHCSzizi⊤. Eigenvalue decomposition is then applied:(12)Cc=UcΛcUc⊤. Here, Cc is a symmetric positive semi-definite scatter matrix, Uc contains the orthonormal eigenvectors, and Λc is the diagonal matrix of eigenvalues. The first principal direction uc, i.e., the eigenvector in Uc corresponding to the largest eigenvalue, represents the dominant feature direction of class c. For each sample xi belonging to class c, an alignment score is computed as(13)siHCS=zi⊤uc. This score measures the consistency between the sample representation and the class-wise principal direction. Clean samples are expected to lie closer to the dominant class subspace and thus exhibit higher alignment scores, while noisy samples tend to deviate from this structure. Following ANNE [5], a two-component Gaussian mixture model is fitted to the class-wise alignment scores to divide DHCS into a clean subset DHCSclean and a noisy subset DHCSnoisy. Since larger alignment scores indicate higher reliability, the component with the larger mean is regarded as the clean component. (b)Low-Confidence Set (LCS) Low-Confidence Set (LCS) For samples in DLCS, which are more likely to contain noisy labels, an adaptive nearest-neighbor strategy adopted in ANNE [5] is employed to exploit local neighborhood structure. This strategy is related to the neighborhood-based selection idea in SSR [37]. Given two samples xi and xj, their cosine similarity is computed using the normalized features zi and zj as(14)sij=zi⊤zj=fϕ(xi)⊤fϕ(xj)fϕ(xi)2fϕ(xj)2. Instead of using a fixed number of neighbors, ANNE [5] introduces an adaptive mechanism. For each sample xi, a similarity threshold ωi is initialized to a high value and iteratively decreased until a minimum number of neighbors Kmin is reached. Specifically, the neighborhood of xi is defined as(15)Ni=j | sij>ωi, j≠i,Ki=Ni If Ki