Computer Vision and Image Understanding最新文献

{"title":"CPNText-DETR: Dynamic contour initialization and normalized loss for accurate arbitrarily shaped text detection","authors":"Yang Wang ,&nbsp;Wanlin Zhou ,&nbsp;Chen Lv ,&nbsp;Hongjun Xia ,&nbsp;Jiakai Zhou","doi":"10.1016/j.cviu.2026.104647","DOIUrl":"10.1016/j.cviu.2026.104647","url":null,"abstract":"<div><div>Recent transformer-based methods for arbitrarily shaped text detection have shown improved performance by predicting text contour points. However, these methods often rely on bounding box centers for text contour points localization, which may not accurately fall within the text area, and use L1 loss for contour point regression, causing imbalances across scales. In this paper, we propose CPNText-DETR, an innovative method that addresses these challenges. Specifically, we introduce a mass-based dynamic point query modeling strategy, which dynamically initializes contour points based on the mass center of the bounding box, aligning them more closely with the actual text shape. Additionally, we present the Contour Point Normalized (CPN) loss, which ensures uniform weighting for contour point regression across texts of varying scales, thus improving regression balance. Furthermore, we introduce the Parallel Enhanced Factorized Self-Attention (PEFSA) module, which captures intra- and inter-textual relationships, enhancing both performance and inference speed. Extensive experiments demonstrate that CPNText-DETR achieves state-of-the-art performance, with F-measures of 89.6%, 89.5%, 90.5% and 80.1% on the Total-Text, CTW1500, ICDAR2015 and Art datasets, respectively. CPNText-DETR not only surpasses existing methods in detection accuracy but also offers faster inference speed, establishing new benchmarks across widely-used datasets. The code is available at <span><span>https://github.com/youngnuaa/CPNText-DETR</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104647"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191768","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"HFINet: Hybrid Feature Integration for enhancing collaborative camouflaged object detection","authors":"Shiyuan Li ,&nbsp;Hongbo Bi ,&nbsp;Disen Mo ,&nbsp;Cong Zhang ,&nbsp;Yue Li","doi":"10.1016/j.cviu.2026.104683","DOIUrl":"10.1016/j.cviu.2026.104683","url":null,"abstract":"<div><div>Collaborative camouflaged object detection aims to identify camouflaged objects with similar characteristics across a group of related images. Detecting multiple camouflaged objects in complex environments is crucial for improving efficiency and accuracy in military operations and security surveillance applications. However, current collaborative camouflaged object detection methods still face several challenges, including insufficient multi-scale feature extraction and integration, limited edge-awareness capabilities, and inadequate inter-group consensus recognition. To solve these challenges, we propose a novel approach: the Hybrid Feature Integration Network (HFINet), designed to revolutionize collaborative camouflaged object detection. HFINet includes three main modules: the cross-layer feature integration (CLFI), the local feature refinement (LFR), and the dual granularity feature fusion (DGFF). By combining feature maps of various resolutions, the CLFI captures multi-scale information from low to high level, the LFR improves the model’s ability to capture details by refining local features using multiple convolution kernels of varying sizes together with an edge-guided attention mechanism to sharpen object boundaries, and the DGFF fuses multi-resolution feature maps to enhance camouflaged object detection, especially to enhance inter-group consensus and synergy performance. Through the experimental results, the HFINet outperforms 35 methods, including 18 existing state-of-the-art camouflaged object detection (COD) methods, 10 state-of-the-art co-salient object detection (CoSOD) methods and 7 state-of-the-art collaborative camouflaged object detection (CoCOD) methods on six regularly evaluated metrics, according to extensive experiments conducted on the CoCOD8K dataset. The code can be obtained at <span><span>https://github.com/LSY20030127/HFINet/tree/master/HFINet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104683"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191767","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"WildMDT-YOLO: A multi-scale wildlife detection model for complex environments","authors":"Chao Li,&nbsp;Youbo Pang,&nbsp;Xianhang Liu,&nbsp;Zhipeng Yu,&nbsp;Minchao Sun","doi":"10.1016/j.cviu.2026.104678","DOIUrl":"10.1016/j.cviu.2026.104678","url":null,"abstract":"<div><div>The deterioration of the global ecological environment and increasing human activities pose severe threats to wildlife survival, making reliable detection methods crucial for wildlife protection and monitoring. However, existing detection methods often encounter inadequate feature extraction in complex settings such as tree occlusion, strong illumination, and low-light environments. To address this challenge, this paper proposes WildMDT-YOLO, an improved YOLOv8n-based wildlife detection method. The model introduces a Multi-scale Focusing Diffusion Network (MFDN) that enhances contextual information across scales through feature focusing and diffusion mechanisms. A novel detection head employs shared convolutions to reduce model parameters while task alignment and interactive feature extraction improve both classification and localization accuracy. The integration of Deformable Convolutional Networks v3 (DCNv3) and Mixed Local Channel Attention (MLCA) mechanism further enhances adaptability to wildlife species with complex shapes and varying scales. Experimental results show that WildMDT-YOLO achieves 92.6% mean average precision (mAP), a 3.1% increase over baseline YOLOv8n, while reducing parameters by 17.1%. Cross-dataset evaluation on Snapshot Serengeti demonstrates robust generalization capability, achieving 89.2% mAP and maintaining 83.6% small object detection performance despite significant domain shift between ecosystems. This model provides an effective tool for improving monitoring efficiency and accuracy in wildlife conservation.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104678"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191769","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Anomaly detection via reconstructing representations from masked compact embedding","authors":"Jiafeng Zhuang ,&nbsp;Xiangzhi Tang ,&nbsp;Junyong Ye ,&nbsp;Xiangning Cui,&nbsp;Yuan Li,&nbsp;Pengyu Feng","doi":"10.1016/j.cviu.2026.104711","DOIUrl":"10.1016/j.cviu.2026.104711","url":null,"abstract":"<div><div>Knowledge distillation methods using T–S (teacher–student) model have recently made remarkable achievements in the field of image anomaly detection (AD), but they still face the problem of overgeneralization which means the student network unexpectedly generalize the expression of normal information to anomalies. We attribute this phenomenon to (1) the propagation of abnormal information which is not strictly restricted and (2) the student network tending to learn the identity mapping from teacher network due to simple training task. To alleviate this problem, we propose the FAE (filter after enhancement) network to enhance input and perform anomaly filtering to get a deep masked embedding, which can not only transfer the training task into a more challenging process of <strong>Re</strong>constructing representation from the <strong>M</strong>asked <strong>C</strong>ompact <strong>E</strong>mbedding (ReMCE), but also effectively limit the propagation of abnormal information in the inference stage. And HarmonyNet is further proposed to optimize the alignment quality of the T–S​ model by learning local details. On the widely used benchmark MVTec AD dataset, ReMCE achieves an excellent image-level anomaly detection AUROC of 99.0%, and is superior to most knowledge distillation-based methods in anomaly location performance. Comprehensive ablation experiments and visualization results confirm our success in addressing overgeneralization.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"266 ","pages":"Article 104711"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147426832","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DFTD-Net: A detail-focused tri-domain network for single-image haze removal","authors":"Yanqiu Wu,&nbsp;Jiawei Yang,&nbsp;Dehong Sun,&nbsp;Lei Wang","doi":"10.1016/j.cviu.2026.104706","DOIUrl":"10.1016/j.cviu.2026.104706","url":null,"abstract":"<div><div>Single-image dehazing poses significant challenges because haze causes detail loss, structural distortion, and uneven contrast, while many existing models either enhance fidelity at high computational cost or improve efficiency at the expense of reconstruction quality. To achieve a better balance, we propose a Detail-focused Tri-domain Network (DFTD-Net) for single-image dehazing. Our approach builds upon a multi-input multi-output U-Net (MIMO-UNet) backbone. To further strengthen feature extraction, we replace the standard convolutions in the residual block (ResBlock) with a detail-enhanced convolution (DEConv), resulting in the detail-enhanced residual block (DEResBlock). Within DEConv, parallel standard and differential convolution branches (e.g., CDC, ADC) inject gradient and structural priors to enrich feature representation. Critically, these branches are re-parameterized into a single standard convolution during inference, introducing no extra parameters or computational overhead. Moreover, we propose a Tri-domain Attention (TDA) that jointly models frequency, spatial, and channel domains to adaptively highlight semantically relevant information. Building on TDA, we design a TDA-guided Feature Fusion Module (TDA-GFFM) to effectively integrate low-level structural and high-level semantic information across the network. Extensive experiments on both real and synthetic datasets demonstrate that DFTD-Net achieves an excellent balance between reconstruction fidelity and efficiency. Comprehensive ablation studies validate each component, and downstream object detection tasks confirm its practical utility.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"266 ","pages":"Article 104706"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147426833","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Robust video anomaly detection via causal feature-guided data augmentation","authors":"Zhaoyi Li ,&nbsp;Chenrui Shi ,&nbsp;Che Sun ,&nbsp;Yuwei Wu","doi":"10.1016/j.cviu.2026.104671","DOIUrl":"10.1016/j.cviu.2026.104671","url":null,"abstract":"<div><div>Data augmentation is a commonly used technique to learn distinguishing features for accurately identifying video anomaly patterns. However, traditional data augmentation methods introduce spurious features-spurious correlations (<em>e.g.</em>, lighting conditions) that lack causal links to anomalies, yielding models that excel on specific test sets but fail in real-world scenarios with varying conditions. In this paper, we propose a novel robust video anomaly detection method by designing a causal feature-guided augmentation technique that strategically enhances the learnability of predictive patterns while suppressing spurious correlations. Specifically, we first disentangle causal features ,<em>i.e.</em>, directly predictive of labels, from spurious features via a causal generative model. We then perturb causal features to enhance their variability in training data, compelling the model to focus on invariant patterns while ignoring spurious correlations. The augmented samples are reconstructed with the refined causal representations, enhancing the model’s discriminative capability. Furthermore, we introduce a systematic evaluation framework with three increasing difficulty levels to assess robustness: seen/unseen variations, cross-dataset generalization, and cross-domain adaptation. This evaluation examines system stability under varying conditions, closely aligning with real-world surveillance deployment requirements. Experimental results validate both the effectiveness and robustness of our method.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104671"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081622","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DepthOcc: Real-time and accurate 3D occupancy prediction via multi-depth fusion and temporal enhancement","authors":"Chenghai Mao ,&nbsp;Yu Qin ,&nbsp;Yiqiang Wu ,&nbsp;Xinghong Zhou ,&nbsp;Bin Zhu ,&nbsp;Jia Liu ,&nbsp;Yan Peng","doi":"10.1016/j.cviu.2026.104670","DOIUrl":"10.1016/j.cviu.2026.104670","url":null,"abstract":"<div><div>3D occupancy prediction, emphasizing occupancy estimation and semantic classification, plays a pivotal role in scene understanding for autonomous driving. Achieving both high accuracy and real-time inference is essential for its practical deployment. However, prevailing lightweight methods are constrained by inadequate depth estimation precision and suboptimal temporal fusion, leading to suboptimal performance. To tackle these significant challenges, we propose DepthOcc, an innovative 3D occupancy prediction method that integrates multi-depth fusion technology and temporal enhancement mechanisms. Specifically, we propose a Multi-Depth Fused View Transformer (MDFVT) that strategically integrates monocular and stereo depth features, substantially enhancing depth estimation precision. Additionally, a Temporal Enhancement Occupancy Head (TEOH) is developed to systematically capture frame-by-frame spatiotemporal correlations through multi-temporal interaction, strengthening the temporal fusion capability of multi-temporal concatenation and thereby significantly enhancing the effectiveness of the feature fusion process. Comprehensive experiments conducted on the Occ3D-nuScenes benchmark demonstrate that DepthOcc achieves a 3.5 mIoU improvement over the baseline while maintaining high inference speed, yielding performance competitive with state-of-the-art approaches.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104670"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147422115","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"UGLF-Net: A parallel architecture for Underwater Global-Local Feature Fusion Network","authors":"Erkang Chen,&nbsp;Wangen Chen,&nbsp;Zhiwei Shen,&nbsp;Zhihui Li,&nbsp;Zhiqi Lin","doi":"10.1016/j.cviu.2026.104704","DOIUrl":"10.1016/j.cviu.2026.104704","url":null,"abstract":"<div><div>Underwater image enhancement is highly challenging due to low contrast, color distortion, blurring caused by light attenuation and scattering. This paper proposes a novel parallel architecture, the Underwater Global-Local Feature Fusion Network (UGLF-Net) for robust image restoration. UGLF-Net consists of the AMFE module for high-quality global feature extraction, the HMCM module with SSM for selective local enhancement and the Swin FAM module for capturing global context. By progressively fusing multi-source features (RGB, grayscale gradients and reduced-dimension data) in a parallel manner, UGLF-Net achieves effective global-local collaborative modeling. Residual connections and Enhanced ECA modules further improve feature representation and training stability, enabling state-of-the-art (SOTA) performance. Experiments on LSUI, EUVP and UIEB datasets show that UGLF-Net outperforms existing methods, including the U-shape Transformer, in PSNR and SSIM. Ablation studies validate the effectiveness of each component. Qualitative results demonstrate superior restoration of vivid colors and fine details. The lightweight design with single-layer SSM and window attention achieves efficient inference (0.009s per image), making it well-suited for real-time enhancement on embedded devices and advancing underwater visual applications.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104704"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147422109","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Frequency domain-based edge sensing for camouflaged object detection","authors":"Bin Ge,&nbsp;Xiaolong Peng,&nbsp;Chenxing Xia,&nbsp;Hailong Chen","doi":"10.1016/j.cviu.2026.104705","DOIUrl":"10.1016/j.cviu.2026.104705","url":null,"abstract":"<div><div>Camouflaged Object Detection (COD), an emerging research direction in computer vision, faces a core challenge: accurately segmenting objects that are naturally or artificially concealed within visually similar backgrounds. In COD tasks, camouflaged objects often exhibit high similarity to their surroundings in terms of texture and color, rendering traditional saliency cues insufficient for reliable target-background discrimination. In contrast, edges serve as structural cues that offer more stable and explicit boundary information, thereby facilitating accurate localization of camouflaged object contours. Motivated by this insight, we propose a Frequency-Guided Edge Encoder (FGEE), which employs a spatial-frequency dual-branch cascaded architecture to enable multi-scale edge modeling and extract more precise and fine-grained edge features. Furthermore, we introduce a Feature Progressive Reinforcement Module (FPRM) that leverages a combination of reverse attention mechanisms and deformable convolutions to suppress foreground distractions and mine structural representations of camouflaged objects for enhanced feature learning. Additionally, we design an Edge-Driven Hierarchical Feature Aggregator (EDHFA) that dynamically integrates contextual information by detecting discrepancies between dual-branch features, generating initial edge contours, and progressively refining edge representations. Extensive experimental results conducted on four widely used COD benchmark datasets demonstrate that the proposed FDESNet surpasses 15 state-of-the-art methods, achieving significant improvements in segmentation performance. The source code is available at <span><span>https://github.com/Pengxiaolong293/FDESNet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104705"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191652","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"ShortNeXt: A novel method for accurate classification of colorectal cancer histopathology images","authors":"Prabal Datta Barua ,&nbsp;Burak Tasci ,&nbsp;Mehmet Baygin ,&nbsp;Sengul Dogan ,&nbsp;Turker Tuncer ,&nbsp;Filippo Molinari ,&nbsp;Salvi Massimo ,&nbsp;U. Rajendra Acharya","doi":"10.1016/j.cviu.2026.104672","DOIUrl":"10.1016/j.cviu.2026.104672","url":null,"abstract":"<div><div>Cancer is a chaotic disease known as the plague of our age and there are many subtypes of the cancer. Cancer is commonly seen disorder and its mortality rate is very high. Therefore, many researchers have worked/studied on the cancer detection and treatment. To contribute cancer studies according to machine learning, we have presented a new generation convolutional neural network (CNN) termed ShortNeXt in this research. The presented ShortNeXt has inspired by ResNet, ConvNeXt and MobileNet architectures to use the advantages these CNNs together. This model, which aims to extract robust feature map using convolution-based residual blocks, is named ShortNeXt because it incorporates more than one shortcut. The ShortNeXt architecture has four main stages and these stages are: (i) an input/stem, (ii) ShortNeXt, (iii) downsampling, and (iv) output. In this CNN architecture, convolution, batch normalization and the Gaussian Error Linear Unit (GELU) activation functions have been utilized. In this aspect, the implementation of the recommended ShortNeXt is simple. The stem stage uses a 4 × 4 sized convolution with stride 4 like ConvNeXt and Swin Transformer and this operation is named patchify operation. Additionally, a 2 × 2 patchify block has been used in the downsampling block. In the ShortNeXt block, an inverted bottleneck has been used, and both 1 × 1 and 3 × 3 convolution blocks are employed in the expansion phase. The output layer has increased the number of filters from 768 to 1280 by using pixel-wise convolution, drawing inspiration from MobileNetV2 and a final feature map with a length of 1280 has been obtained by deploying global average pooling (GAP). In the classification phase, fully connected and softmax operators have been used.</div><div>To get comparative results about to the recommended ShortNeXt, a publicly available histopathological image dataset has been used and this dataset contains nine classes, and the proposed ShortNeXt has achieved 97.82% and 97.86% validation and test accuracy, respectively. The obtained results and findings openly showcases that ShortNeXt is an effective deep learning method for histopathological image classification for cancer detection/classification.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104672"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191772","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}