Video anomaly detection is crucial for analyzing surveillance videos and plays a significant role in maintaining public safety. Recent advances in weakly supervised methods, utilizing video-level labels, have improved performance based on techniques like multi-instance learning and temporal modeling. Furthermore, memory banks demonstrate great potential in unsupervised anomaly detection, prompting their integration into weakly supervised setups. However, these methods depend on the Top-k selection mechanism to update the prototypes within memory banks, which has limitations such as overlooking valuable prototypes, leading to a biased updating process, and requiring hyperparameters. To tackle these challenges, we introduce a novel video anomaly detection model, FADMB (Fully Attention-based Dual Memory Bank network), which replaces the Top-k selection mechanism with an innovative attention-based prototype updating paradigm to obtain a more comprehensive and robust memory bank. Additionally, we design a Hybrid Encoder that encodes local and global temporal information to produce superior video representations. Extensive experiments demonstrate the superiority of FADMB, achieving 85.79% AUC on UCF-Crime dataset and 83.29% AP on XD-Violence dataset.

Visible-infrared person re-identification (VI-ReID) faces large challenges due to the cross-modality gap and intraclass variations. Existing methods mainly focus on mitigating the cross-modality discrepancy by optimizing the embedding space of high-level features. However, two critical issues are often overlooked: low-level features suffer from larger modality gaps and redundant channels exhibit limited identity discrimination. These limitations hinder the overall discriminative performance of the model. To address the above issues, we propose a novel cross-modality self-distillation (CMSD) network, which consists of two key components: cross-layer semantic distillation (CSD) and intralayer alignment distillation (IAD). These two modules are mutually beneficial to jointly explore cross-layer modality discrepancy and intralayer channel discrepancy. Specifically, CSD distills modality-shared semantics from high-level features to guide low-level features in learning more reliable modality-invariant representations. Furthermore, IAD decomposes feature maps into salient and redundant channels while transferring identity-related information from salient to redundant channels, thereby enhancing overall discrimination of modality-invariant features. Our extensive experiments on the SYSU-MM01, RegDB, and LLCM datasets demonstrate that the components of our proposed CMSD mutually enhance each other, leading to superior performance compared to state-of-the-art (SOTA) methods.

Vessel re-identification (ReID) plays a critical role in maritime surveillance by matching vessels across different camera views. Compared with person or vehicle ReID, vessel ReID faces unique challenges due to subtle interclass differences and large intraclass variations caused by viewpoint changes. These issues are further exacerbated by the highly similar appearances of vessels and the lack of fine-grained identity cues commonly found in other ReID tasks. To address these challenges, we propose a spatial-channel fusion network (SCF-Net), a dual-branch deep framework that integrates a spatial-channel fusion (SCF) module and a feature refinement and alignment (FRA) module. The SCF module captures interdependent relationships between spatial and channel dimensions, enabling the network to emphasize discriminative regions while suppressing irrelevant background information. The FRA module refines high-dimensional embeddings into a compact representation and enforces intraclass similarity via a learnable multilayer perceptron (MLP) and a supervised mean squared error (MSE) loss. By jointly optimizing the two branches and the FRA output, SCF-Net effectively learns both interclass discrimination and intraclass compactness. Extensive experiments demonstrate that SCF-Net achieves competitive performance on public vessel ReID benchmarks, highlighting its effectiveness in handling subtle interclass differences and large intraclass variations.

Inverse adversarial training leverages high-confidence predictions to stabilize robust learning, yet we uncover a critical paradox: high confidence often stems from overfitting to non-causal background correlations rather than intrinsic object semantics. Our investigation reveals that visual context functions as a dual-natured signal, serving as either a necessary supportive prior or a spurious confounder. This insight renders existing blind suppression strategies flawed, as they inevitably lead to severe Feature Loss. To resolve this, we propose High-Confidence Causally Aligned Training (HICAT), a unified framework that establishes a Semantic Equilibrium. Operating on a “Measure-Debias-Align” pipeline, HICAT integrates a Learnable Background-Bias Estimator (LBBE) to adaptively diagnose context utility. Guided by this diagnosis, an Adaptive Debiasing mechanism performs surgical logit rectification, complemented by a geometrically grounded Foreground Logit Orthogonal Enhancement (FLOE) loss to enforce rigorous feature disentanglement. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet-1K demonstrate that HICAT consistently improves over matched baselines across diverse architectures (CNNs and ViTs) while significantly reducing the robust generalization gap.

Recent studies have shown that Deep Neural Networks (DNNs) are susceptible to adversarial attacks, with frequency-domain analysis underscoring the significance of high-frequency components in influencing model predictions. Conversely, targeting low-frequency components has been effective in enhancing attack transferability on black-box models. In this study, we introduce a frequency decomposition-based feature mixing method to exploit these frequency characteristics in both clean and adversarial samples. Our findings suggest that incorporating features of clean samples into adversarial features extracted from adversarial examples is more effective in attacking normally-trained models, while combining clean features with the adversarial features extracted from low-frequency parts decomposed from the adversarial samples yields better results in attacking defense models. However, a conflict issue arises when these two mixing approaches are employed simultaneously. To tackle the issue, we propose a cross-frequency meta-optimization approach comprising the meta-train step, meta-test step, and final update. In the meta-train step, we leverage the low-frequency components of adversarial samples to boost the transferability of attacks against defense models. Meanwhile, in the meta-test step, we utilize adversarial samples to stabilize gradients, thereby enhancing the attack's transferability against normally trained models. For the final update, we update the adversarial sample based on the gradients obtained from both meta-train and meta-test steps. Our proposed method is evaluated through extensive experiments on the ImageNet-Compatible dataset, affirming its effectiveness in improving the transferability of attacks on both normally-trained CNNs and defense models. The source code is available at https://github.com/WJJLL/MetaSSA.

Occlusion presents two critical challenges for person re-identification (Re-ID): feature interference and information loss. While existing efforts have explored occlusion-aware data augmentation and feature reconstruction to mitigate these issues, the former often fails to address erroneous matches caused by similar occlusion patterns and background distractions, whereas the latter typically introduces significant computational overhead. To overcome these limitations, we propose a Consistent Occlusion and Prompt Enhancement (COPE) network. COPE incorporates a Cross-Identity Consistent Occlusion (CICO) module that applies identical occlusions across different identities and encourages feature similarity in the same occluded regions across different identities to reduce occlusion feature interference. A Prompt Background Filling (PBF) module leverages vision-language alignment to generate foreground heatmaps and performs random background filling, enhancing feature robustness under varying backgrounds. Additionally, a lightweight Prompt Similarity Scoring (PSS) module refines retrieval similarity by utilizing prompt-guided reliability scores. Extensive experiments on both occluded and holistic Re-ID benchmarks demonstrate that COPE consistently outperforms existing methods. Notably, it achieves 82.4% Rank-1 accuracy and 76.4% mAP on the challenging Occluded-Duke dataset.

Despite recent advancements in font generation, practitioners still grapple with a laborious trial-and-error workflow. To streamline this, we propose OneFont, an end-to-end framework that interprets user intents via free-form dialogue, seamlessly integrating both glyph synthesis and refinement modules. We introduce the Font with Thought (FwT) paradigm, reframing font design as a reasoning task where the model plans actions and articulates design rationales. OneFont’s core planner is trained via a two-stage regimen to master this paradigm. First, we instill reasoning abilities via Supervised Fine-Tuning (SFT) on a new, comprehensive benchmark of 1,500 font families we built. Second, we refine the model's policy with a novel reinforcement learning algorithm, Group Relative Policy Optimization (GRPO), guided by a hybrid reward that assesses visual fidelity, rationale coherence, and transformation correctness. Extensive experiments show OneFont significantly surpasses existing methods in design quality and stroke precision across diverse scripts, validated on our new benchmark. We will release our dataset, code, and models.

Accurate and automated tumor segmentation in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is vital for breast cancer diagnosis and treatment. However, the substantial heterogeneity of breast cancer, along with the varied tumor sizes, shapes, and appearances, coupled with artifacts in DCE-MRI data, pose significant challenges for accurate segmentation. Furthermore, current approaches encounter difficulties in cross-dataset scenarios with different image acquiring protocols and devices, due to ineffective representation of essential tumor features. To address these challenges, we propose a Multi-level Temporal-Spatial Contrastive Learning Network (MTSCL-Net) for breast tumor segmentation in DCE-MRI. Our method introduces a novel multi-level temporal-spatial contrastive loss to enhance feature representation from multiple layers and across temporal sequences. Additionally, we design a feature-sharing encoding structure with tumor-invariant feature perception, reducing parameters while maintaining consistent spatial feature extraction. A temporal fusion module integrates sequence features, further reducing parameter count and complexity. Extensive experiments on two public datasets demonstrate the superiority of our approach over recent state-of-the-art methods. To explore generalization across different centers, we trained our method on a public dataset (DUKE) and tested it on another public dataset (YUN, collected from seven centers) and two private datasets. The results verified the robustness and effectiveness of our approach in addressing both within-domain and cross-domain challenges.

Breast-conserving surgery (BCS) is the standard care for early breast cancer. However, positive surgical margins lead to high recurrence and require reoperation, remaining a clinical challenge. This study aims to develop an ensemble model to preoperatively predict positive margins using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). A total of 887 patients from three Chinese medical centers were included. Patients from the Yunnan (581 patients) and Fujian (51 patients) provinces were used for model development, with external testing on 188 patients from Guangdong and prospective testing on 67 patients from Yunnan. Three base models were developed: the clinical model based on 32 features, the deep learning (DL) model using ResNet-34, and the radiomics model utilizing radiomics features. The ensemble model combined predictions from these three base models using weighted voting. Performance metrics (accuracy, AUC, sensitivity, specificity) were computed, and prognostic analyses were conducted with a public dataset. The ensemble model outperformed each of the three base models, with the AUC of 0.931, accuracy of 0.872, sensitivity of 0.871, and specificity of 0.880 in the validation set. Robustness was further confirmed in independent test cohorts, yielding AUCs of 0.762 and 0.861. Furthermore, patients identified at high-risk by the model exhibited poor recurrence-free survival (RFS, HR = 8.117, p < 0.0001) and overall survival (OS, HR = 5.748, p < 0.0001), underscoring its prognostic value. The ensemble model based on DCE-MRI effectively predicts positive margins and prognosis for BCS patients, facilitating the clinical decision-making on the surgery. Trial registration This study was approved by the Ethics Committee of the Yunnan Cancer Hospital (KYLX2023-134), Xiang’an Hospital of Xiamen University (XDYX202305K26), and Cancer Hospital of Shantou University Medical College (2024059). The study was also registered with the Chinese Clinical Trial Registry (ChiCTR2400083298, Registered 19 April 2024, https://www.chictr.org.cn/showproj.html?

Feature selection is a vital preprocessing step aimed at identifying a subset of the most relevant features from high-dimensional data to enhance model performance and reduce computational complexity. Differential Evolution (DE) algorithms have been extensively applied to this task by iteratively optimizing the selection probabilities or weights of features. However, many existing DE-based approaches suffer from premature convergence and local optima entrapment due to an insufficient balance between global exploration and local exploitation. To address these challenges, we propose a niching-based two-stage mutation DE algorithm for feature selection. Firstly, the mutual information is utilized to initialize the population and reduce the number of features. Then, an improved two-stage mutation operator is employed to balance the algorithm's exploitation and exploration. Additionally, duplicate individuals generated during the evolutionary process are replaced using a one-bit evolutionary mutation to aid population evolution. Experimental results on 16 benchmark datasets demonstrate that the proposed method achieves superior classification accuracy compared to several state-of-the-art approaches, validating its effectiveness and robustness in diverse feature selection scenarios.

Ensuring the robustness of deep neural networks against adversarial attacks remains a fundamental challenge in computer vision. While adversarial training (AT) has emerged as a promising defense strategy, our analysis reveals a critical limitation: AT-trained models exhibit a bias toward low-frequency features while neglecting high-frequency components. This bias is particularly concerning as each frequency component carries distinct and crucial information: low-frequency features encode fundamental structural patterns, while high-frequency features capture intricate details and textures. To address this limitation, we propose High-Frequency Feature Disentanglement and Recalibration (HFDR), a novel module that strategically separates and recalibrates frequency-specific features to capture latent semantic cues. We further introduce frequency attention regularization to harmonize feature extraction across the frequency spectrum and mitigate the inherent low-frequency bias of AT. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet-1K demonstrate that HFDR consistently enhances adversarial robustness. It achieves a 2.89 % gain on CIFAR-100 with WRN34-10, and improves robustness by 3.09 % on ImageNet-1K, with a 4.89 % gain on ViT-B against AutoAttack. These results highlight the method's adaptability to both convolutional and transformer-based architectures. Code is available at https://github.com/KejiaZhang-Robust/HFDR.

Despite the remarkable progress of deep neural networks (DNNs) in various visual tasks, their vulnerability to adversarial examples raises significant security concerns. Recent adversarial training methods leverage inverse adversarial attacks to generate high-confidence examples, aiming to align adversarial distributions with high-confidence class regions. However, our investigation reveals that under inverse adversarial attacks, high-confidence outputs are influenced by biased feature activations, causing models to rely on background features that lack a causal relationship with the labels. This spurious correlation bias leads to overfitting irrelevant background features during adversarial training, thereby degrading the model's robust performance and generalization capabilities. To address this issue, we propose Debiased High-Confidence Adversarial Training (DHAT), a novel approach that aligns adversarial logits with debiased high-confidence logits and restores proper attention by enhancing foreground logit orthogonality. Extensive experiments demonstrate that DHAT achieves state-of-the-art robustness on both CIFAR and ImageNet-1K benchmarks, while significantly improving generalization by mitigating the feature bias inherent in inverse adversarial training approaches. Code is available at https://github.com/KejiaZhang-Robust/DHAT.

Natural Language-Guided Drones (NLGD) offer a novel and flexible interaction paradigm for tasks such as target matching and navigation. However, the wide field of view and complex compositional semantic relationships inherent in drone scenarios place greater demands on visual language understanding. First, mainstream Vision-Language Models (VLMs) primarily focus on global feature alignment and lack fine-grained semantic understanding. Second, existing hierarchical semantic modeling methods rely on precise entity partitioning and strict containment relationship constraints, which limits their effectiveness in complex drone environments. To address these challenges, we propose the Hierarchical Cross-Granularity Contrastive and Matching learning (HCCM) framework, comprising two core components: 1) Region-Global Image-Text Contrastive Learning (RG-ITC). Avoiding precise scene entity partitioning, RG-ITC models hierarchical local-to-global cross-modal semantics by contrasting local visual regions with global text semantics, and vice versa. 2) Region-Global Image-Text Matching Learning (RG-ITM). Instead of relying on strict relationship constraints, this component evaluates local semantic consistency within global cross-modal representations, improving the comprehension of complex compositional semantics. Furthermore, drone scenario textual descriptions are often incomplete or ambiguous, destabilizing global semantic alignment. To mitigate this, HCCM incorporates a Momentum Contrast and Momentum Distillation (MCD) mechanism, enhancing alignment robustness. Extensive experiments on the GeoText-1652 benchmark demonstrate HCCM significantly outperforms existing methods, achieving state-of-the-art Recall@1 scores of 28.8% (image retrieval) and 14.7% (text retrieval). Moreover, HCCM exhibits strong zero-shot generalization on the unseen ERA dataset, achieving 39.93% mean recall (mR), surpassing evaluated fine-tuned models. These results highlight the effectiveness and robustness of HCCM across diverse scenarios. Our implementation is available at https://github.com/rhao-hur/HCCM.

Video tasks play an important role in multimedia fields. In various video tasks, such as video camouflaged/salient object detection (VCOD/VSOD), motion and context information are two important aspects. Despite the fact that many existing works have already achieved promising results in VCOD and VSOD tasks, they still have limitations when it comes to leveraging motion and context information. In this paper, we propose a new disentangled perspective to treat motion and context information in VCOD and VSOD tasks. Our proposed model can respectively utilize context and motion information in ContextNet and MotionNet, without conflicting with each other as there can be biases between these two types of information in certain circumstances. Moreover, we further explore how to apply disentangled perspective in the self-supervised manner, which can reduce annotation costs. Specifically, we first design a self-supervised adaptive frame routing mechanism to determine whether each video frame belongs to ContextNet or MotionNet. Then we design a cross-supervision for ContextNet and MotionNet to train these two segmentation networks in self-supervised mechanism. In experiments, our proposed self-supervised disentangled model consistently outperforms state-of-the-art unsupervised methods on VCOD and VSOD datasets.

Weakly supervised video anomaly detection plays a pivotal role in widely deployed surveillance systems. Most existing methods are based on the multi-instance learning paradigm, determining the predicted label of a video based on segments with higher prediction scores. During training, the model predominantly focuses on the segment with the highest anomaly score or the top-k highest-scored segments, neglecting the other segments. This bias towards certain segment features during training results in missed and false detections of anomaly segments, subsequently impacting the performance of video anomaly detection. In this paper, we introduce a contrastive loss strategy to uncover easily overlooked normal and abnormal segments, enhancing the distinction between normal and abnormal segments through contrastive loss. Additionally, we propose a multi-scale feature fusion approach to learn features from different scales of videos and integrate them into a more comprehensive feature representation to accommodate the diversity of anomaly events. Experimental results on the UCF-Crime and XD-Violence datasets validate the efficacy of our proposed method.

Tongue diagnosis in Traditional Chinese Medicine (TCM) is a crucial diagnostic method that can an individual's health status. Traditional methods for identifying tooth-marked tongues are subjective inconsistent because they rely on practitioner experience. We propose a novel fully automated W Supervised method using Vision transformer and Multiple instance learning (WSVM) for tongue extraction and tooth-marked tongue recognition. Our approach first accurately detects and extracts the tongue from clinical images, removing any irrelevant background information. Then, we implement an end-to weakly supervised object detection method. We utilize Vision Transformer (ViT) to process tongue images patches and employ multiple instance loss to identify tooth-marked regions with only image-level annotations. WSVM achieves high accuracy in tooth-marked tongue classification and tooth-marked tongue detection. Visualization experiments further demonstrate its effectiveness in pinpointing these regions. This automated approach enhances the objectivity and accuracy of tooth-marked tongue diagnosis. It provides significant clinical value by assisting TCM practitioners in making precise diagnoses and treatment recommendations. Code is available at https://github.com/yc-zh/WSVM.

This paper aims to enhance the transferability of adversarial samples in targeted attacks, where attack success rates remain comparatively low. To achieve this objective, we propose two distinct techniques for improving the targeted transferability from the loss and feature aspects. First, in previous approaches, logit calibrations used in targeted attacks primarily focus on the logit margin between the targeted class and the untargeted classes among samples, neglecting the standard deviation of the logit. In contrast, we introduce a new normalized logit calibration method that jointly considers the logit margin and the standard deviation of logits. This approach effectively calibrates the logits, enhancing the targeted transferability. Second, previous studies have demonstrated that mixing the features of clean samples during optimization can significantly increase transferability. Building upon this, we further investigate a truncated feature mixing method to reduce the impact of the source training model, resulting in additional improvements. The truncated feature is determined by removing the Rank-1 feature associated with the largest singular value decomposed from the high-level convolutional layers of the clean sample. Extensive experiments conducted on the ImageNet-Compatible, CIFAR-10 and ImageNet-1k datasets demonstrate the individual and mutual benefits of our proposed two components, which outperform the state-of-the-art methods by a large margin in black-box targeted attacks.

Current out-of-distribution (OOD) detection methods typically assume balanced in-distribution (ID) data, while most real-world data follow a long-tailed distribution. Previous approaches to long-tailed OOD detection often involve balancing the ID data by reducing the semantics of head classes. However, this reduction can severely affect the classification accuracy of ID data. The main challenge of this task lies in the severe lack of features for tail classes, leading to confusion with OOD data. To tackle this issue, we introduce a novel Prioritizing Attention to Tail (PATT) method using augmentation instead of reduction. Our main intuition involves using a mixture of von Mises-Fisher (vMF) distributions to model the ID data and a temperature scaling module to boost the confidence of ID data. This enables us to generate infinite contrastive pairs, implicitly enhancing the semantics of ID classes while promoting differentiation between ID and OOD data. To further strengthen the detection of OOD data without compromising the classification performance of ID data, we propose feature calibration during the inference phase. By extracting an attention weight from the training set that prioritizes the tail classes and reduces the confidence in OOD data, we improve the OOD detection capability. Extensive experiments verified that our method outperforms the current state-of-the-art methods on various benchmarks.

Unsupervised visible-infrared person re-identification (USVI-ReID) aims to match a person across two modalities without annotations. Current research primarily addresses the modality gap by establishing cross-modality correspondences through matching algorithms and utilizing memory banks for contrastive learning. However, the inherent noise in pseudo labels and neglect of hard samples often limit the efficacy of cross-modality learning. In this article, we propose a dual-modality-shared learning and label refinement (DLLR) algorithm for USVI-ReID. First, we leverage a cluster similarity matching (CSM) module and a cluster relationship-based label refinement (CRLR) algorithm to create and refine pseudo labels. Then, we adopt a weighted modality-shared memory (WMM) to construct memory banks by jointly considering sample distribution and feature differences, thereby enhancing the effectiveness of cross-modality learning. Extensive experiments on three publicly available datasets validate the effectiveness of our proposed method, which outperforms state-of-the-art methods. The code is available at https://github.com/CharRic/DLLR.