Weakly Supervised Semantic Segmentation (WSSS) aims to learn pixel-level semantic predictions using only image-level annotations. However, due to the absence of precise spatial supervision, the generated Class Activation Maps (CAMs) often highlight only the most discriminative regions of objects, resulting in incomplete object coverage and unstable cross-layer semantic responses. To address these challenges, we propose a token-level contrastive learning based framework for WSSS, which improves CAM localization quality by enhancing feature representation and enforcing cross-layer structural consistency. Specifically, we first introduce a multi-subspace token-level contrastive module, which decouples feature representations through a shared semantic backbone and multiple projection subspaces, thereby increasing the diversity and discriminability of the embedding space. Furthermore, we propose a cross-layer CAM structural alignment module that jointly constrains both the response intensity and spatial structural relationships of CAMs across different Transformer layers, leading to more stable semantic localization and improved spatial consistency of object regions. Extensive experiments on the PASCAL VOC 2012 and MS COCO 2014 benchmarks demonstrate that the proposed method consistently improves segmentation performance under an end-to-end training framework. In particular, it achieves 71.8% (val) and 72.3% (test) mIoU on VOC 2012, and 42.6% mIoU on the COCO 2014 validation set. Further ablation studies validate the effectiveness of each component. Overall, our method significantly enhances the completeness and structural stability of CAMs, providing an effective solution for representation learning and structural modeling in WSSS.

Few-shot fine-grained image classification faces challenges due to high inter-class feature similarity, significant intra-class variations, and limited data, leading to insufficient model generalization. To address the challenges of category aggregation and separation in few-shot fine-grained scenarios, this paper proposes a classification network enhanced by cross-sample feature interaction. The cross-sample feature interaction enhancement process includes three core components: the Cross-domain Feature Attention Network (CFAN), which enhances feature consistency between support and query sets through channel and spatial attention mechanisms, focusing on capturing critical detail regions; the Global Dependency Augmentation Module (GDAM), which explicitly models dependencies between distant pixels and integrates local and global information; and the Cross-class Interaction Module (CCIM), which uses a bidirectional self-attention mechanism to align local features of query and support samples through complementary and interactive cross-sample features. Additionally, dynamic generation of inter-class feature contrast relationships improves inter-class feature discrimination. Experimental results demonstrate that the proposed method achieves superior performance compared to mainstream methods on multiple public few-shot fine-grained classification datasets, with particularly remarkable results under extremely few-shot conditions. The proposed method significantly improves the accuracy and robustness of few-shot fine-grained classification through an efficient feature interaction mechanism.

Incremental learning often encounter challenges such as overfitting to new data and catastrophic forgetting of old data. Existing methods can effectively extend the model for new tasks while freezing the parameters of the old model, but ignore the necessity of structural efficiency to lead to the feature difference between modules and the class misalignment due to evolving class distributions. To address these issues, we propose scalable class-incremental learning based on parametric neural collapse (SCL-PNC) that enables demand-driven, minimal-cost backbone expansion by adapt-layer and refines the static into a dynamic parametric Equiangular Tight Frame (ETF) framework according to incremental class. This method can efficiently handle the model expansion question with the increasing number of categories in real-world scenarios. Additionally, to counteract feature drift in serial expansion models, the parallel expansion framework is presented with a knowledge distillation algorithm to align features across expansion modules. Therefore, SCL-PNC can not only design a dynamic and extensible ETF classifier to address class misalignment due to evolving class distributions, but also ensure feature consistency by an adapt-layer with knowledge distillation between extended modules. By leveraging neural collapse, SCL-PNC induces the convergence of the incremental expansion model through a structured combination of the expandable backbone, adapt-layer, and the parametric ETF classifier. Experiments on standard benchmarks demonstrate the effectiveness and efficiency of our proposed method. Our code is available at https://github.com/zhangchuangxin71-cyber/dynamic_ ETF2. Keywords: Class incremental learning; Catastrophic forgetting; Neural collapse;Knowledge distillation; Expanded model.

Few-shot learning can find the latent structure information between the support and query data by the similarity metric of meta-learning to construct the discriminative model for recognizing the new categories with the less labeled samples. Most existing methods attempt to model the similarity relationships among samples within meta-tasks for achieving this goal. However, the relationship assessment among samples from distinct meta-tasks is challenging due to the differing metric relationships inherent to each respective meta-task. To address this issue, high-order structure-preserving graph neural network (HOSP-GNN) can deeply explore the rich samples structure of the different meta-tasks to predict the label of the queried data based on the graph. HOSP-GNN can mine high-order structures to enhance their relevance by triple samples. In addition, it can also generate the updating rule of the different-order structures for node and edge representation optimization under manifold constraints. Notably, HOSP-GNN eliminates the need for retraining the learning model to recognize new classes, thanks to its high generalization of high-order structure that ensures model adaptability. The experiments demonstrate that HOSP-GNN outperforms state-of-the-art methods in four benchmark datasets, as well as on a self-built dataset focusing on endangered animals. The available code is https://github.com/yangfeifei02/HOSP.

To tackle the challenges of capturing discriminative features and effectively leveraging multi-granularity information in fine-grained image classification, this paper proposes an Interactive Modeling Network with Feature Visibility Optimization (IMN-FVO), based on a visibility-guided and interactive learning strategy. IMN-FVO enhances the utilization of multi-scale features and improves the extraction of subtle discriminative details through comprehensive feature mining. It comprises three key modules: (1) the Feature Visibility Mining Unit, which enhances salient features and suppresses redundant ones to improve feature representation and discrimination; (2) the Strip Convolution Optimization Module, which focuses on precise localization by filtering out irrelevant information; and (3) the Interactive Multi-layer Perceptron Module, which models multi-level feature interactions to enrich semantic representation and fusion. The entire framework is trained end-to-end, without requiring bounding box annotations or multi-stage processing. Experiments on CUB-200–2011, Stanford Cars, and FGVC-Aircraft show that IMN-FVO outperforms existing state-of-the-art methods, demonstrating strong effectiveness and generalization.

Defect detection in printed materials is a critical yet challenging task in industrial quality control due to the small size of defects, low contrast against complex backgrounds, and the high-speed requirements of production lines. Existing detectors often struggle to capture subtle features or suppress redundant noise, leading to missed or false detections. To address these limitations, we propose HA-YOLOv8s, an enhanced YOLOv8s model built upon a hybrid-domain attention mechanism. The architecture integrates three tailored modules: (1) a Deep Multi-scale Selective Fusion (DMSF) module that combines depthwise separable residual blocks with multi-scale paths to capture both local and global semantics; (2) a Separated Enhancement Aggregation Module (SEAM) that applies spatial and channel attention to highlight critical features and suppress background interference; and (3) a Separable Pyramid Attention Detection Head (SPAD-Head) that introduces atrous convolutions, SE attention, and an additional 160 × 160 detection layer to improve scale adaptability. These modules collectively improve the model’s capability to detect small, complex defects. Experiments on a printed defect dataset and VisDrone2019 demonstrate that HA-YOLOv8s achieves mAP scores of 98.3

The existing fine-grained image classification algorithms focus on mining low-level details to emphasize subtle differences between subcategories. However, by integrating low-level details with high-level semantics, it is easy to introduce some redundant features (such as blurred regions and background noise), which may hinder the network's understanding of object parts. Therefore, we propose a metric learning and feature reconstruction network to efficiently fuse multi-level features and reduce redundant information. First, the network employs a convolution alignment enhancement module to refine shallow details and uses large kernel convolution to increase the receptive field, capturing a wider range of context information. Second, the network utilizes a metric learning framework to deeply explore inter-layer channel difference information and complementary fusion of these difference features, ensuring the diversity of the fused feature information. Finally, to reduce the interference of redundant information, spatial and channel reconstruction modules are introduced to suppress spatial redundancy and channel redundancy. The proposed method achieves great performance on the three benchmark datasets of CUB-200-2011, FGVC-Aircraft, and Stanford Cars, with accuracy rates of 90.2%, 93.7%, and 95.1%, respectively. The experimental results show that the model is better than the current mainstream fine-grained image classification methods. The classification results are more accurate.

With the widespread adoption of ultra-high-resolution (UHR) images in fields such as photography and social media, traditional image retouching methods face challenges when handling UHR portrait images, including low efficiency, poor consistency in retouching, and limited detail restoration capabilities. To systematically address all of these issues, this paper proposes an adaptive UHR portrait retouching method based on regional filters. This approach mimics the strategy employed by professional human retouchers, who perform localized adjustments based on distinct regional features. Specifically, we introduce a Feature Pyramid Network structure to fuse multi-scale features and integrate a Spatial Prior Fit Module (SPFM) module at the neck stage to enhance both spatial modeling and channel representation. Furthermore, we design a Convolutional Attention Feedforward Block (CFBK) that captures spatial and channel contextual information through local convolutional attention and a lightweight feedforward network. In addition, a Selective Enhancement Prediction Module (SEPM) is developed to selectively enhance regional features and adaptively predict filter parameters such as contrast, saturation, and brightness, enabling efficient and fine-grained portrait retouching. Extensive experiments conducted on the MIT-Adobe FiveK, PPR10K, and our proprietary ALSPP5K datasets demonstrate that the proposed method outperforms existing state-of-the-art approaches in terms of objective metrics. For instance, on PPR10K-a, it achieves improvements of up to 5.75 dB in PSNR and 0.021 in SSIM. These results verify that our method not only maintains high-quality retouching effects but also significantly improves the aesthetic quality and regional detail consistency of images, thus confirming the effectiveness of the regional filter adaptation mechanism in portrait retouching tasks.

The objective of this paper is to address the challenges associated with acquiring defect samples and detecting small defects in print defect detection. We propose a methodology for generating defect samples using the pix2pix-HDNet network, which first generates coarse defects in designated areas of defect-free samples and then adjusts their brightness for visual consistency with the background, resulting in realistic defect samples. Additionally, we improve YOLOv7 and propose a print defect detection algorithm called DF-YOLOv7, which comprises two innovative modules: the Detail Extraction Module (DEM) and the Feature Enhancement Module with Omni-dimensional Dynamic Convolution (FEM-ODConv). DEM enhances the network’s ability to capture detailed information and refine the edge information of defects, while FEM-ODConv improves the extraction of contextual and semantic information, enhancing the weak features of small defects and effectively suppressing background interference. We collected three common print defects, namely ink blot, black point, and missing print, and expanded the defect samples using pix2pix-HDNet. Training DF-YOLOv7 with the expanded dataset achieved an accuracy of 92.1

Small object detection is a critical research area in computer vision, with broad applications in industrial defect detection and satellite remote sensing. In printing defect detection, defects on printed materials are often small, weak in detail, and low in contrast. While mainstream deep learning-based object detection algorithms perform well on conventional objects, they face significant challenges in extracting features for printing defects. Additionally, the small size and blurred visual characteristics of defects make detection results highly susceptible to background interference, leading to a high false positive rate. To address these issues, this paper proposes a progressive printing defect detection method based on contextual information (PCINet). Specifically designed for small defects with unclear visual features, PCINet enhances defect feature representation by reconfiguring the backbone network during the feature extraction phase, thereby improving detection performance. A global semantic reconstruction module is introduced to progressively explore the contextual relationships between defect targets and their surrounding environment. This module includes a global semantic awareness unit, which expands the receptive field and enriches regions of interest, and a regional interaction-assisted reconstruction unit, which refines defect edges and suppresses redundant background interference. Experimental results demonstrate that the proposed method performs well on the printing defect detection dataset, the Printing Defect Dataset 2, and the DOTA-V1.0 dataset. It significantly reduces the false positive rate and exhibits strong robustness in detecting defects under complex backgrounds and low contrast. Furthermore, PCINet shows good generalization capabilities in other small object detection tasks, underscoring its broad application potential.

In real-world scenarios, images captured by devices such as smartphones and cameras often contain diverse content and various types of distortions. Compared to synthetic distortions, real images may encounter more complex, non-uniform artifacts such as ghosting, overexposure, and motion blur, which pose significant challenges for quality assessment. To address the differences between real and synthetic distortions, this paper proposes a no-reference image quality assessment (NR-IQA) algorithm based on a self-attention Transformer encoder, aimed at effectively handling the complexity and diversity of distortions in real-world images. Unlike artificially synthesized distortions, real images captured by mobile phones or cameras are often affected by non-uniform artifacts like ghosting, overexposure, and motion blur, making quality evaluation more challenging. To tackle these issues, the proposed approach exclusively employs the encoder part of a multi-head self-attention mechanism Transformer architecture. Distinct from the original Transformer structure, only the encoder is used for image feature modeling, with the decoder removed to reduce model complexity and better suit the regression nature of NR-IQA tasks. The encoder employs multi-head self-attention to capture long-range dependencies between spatial features, enhancing feature representation and enabling comprehensive modeling of both local and global distortion characteristics. The entire framework consists of a simplified ResNet50 backbone (with Block4 removed to better preserve local details), a multi-head self-attention Transformer encoder, and a multi-layer perceptron for quality score prediction. Additionally, the model introduces positional embeddings and learnable additional tokens to strengthen its ability to handle various distortion types. Extensive experiments conducted on multiple real-world distortion datasets, along with data augmentation strategies, demonstrate the superior generalization performance of the proposed method across different datasets. Compared to existing state-of-the-art approaches, this algorithm exhibits higher accuracy and robustness, highlighting its practical value and broad application prospects in the field of image quality assessment.

Addressing the safety hazards arising from inadequate clearance between tree growth and power facilities, this paper innovatively proposes an online monitoring and early warning system for distribution networks based on advanced image processing technology. The system integrates three core functions: automatic tree species identification, precise detection of safety distances between tree crowns and live equipment, and real-time foreign object contact warnings. By deploying maintenance-free online monitoring terminals, the system can monitor tree growth around the distribution network around the clock, continuously. It utilizes efficient backend algorithms to intelligently analyse the collected image data, promptly detecting and warning of potential tree-obstacle hazards. Compared to traditional manual inspections and UAV inspections, this system not only significantly improves monitoring accuracy and real-time performance but also overcomes the limitations of UAV inspections, such as high costs and poor real-time performance. The solution proposed in this paper is expected to fundamentally enhance the safe operation level of distribution networks and effectively reduce power accidents such as tripping and short circuits caused by tree obstacles. Future plans include further optimizing system functions to promote the wider application and promotion of the technology.

The long-tail distribution is characterized by a large number of samples in a few categories (head classes) and a scarcity of samples in most categories (tail classes). This inherent class imbalance significantly degrades the performance of conventional classification models, particularly on tail classes. To tackle this challenge, we propose a Multi-Expert Dynamic Gating and Feature Decoupling Classification Algorithm based on Uniform Enhanced Sampling. The proposed method integrates multi-expert learning with data augmentation and enhances tail classes performance by jointly optimizing the loss function and the expert assignment network. Specifically, a uniform enhanced sampling strategy is introduced to augment tail classes samples and increase their sampling frequency through resampling. During the feature learning stage, the shared layers of a convolutional network extract general features, while multiple expert models are trained independently. A feature decoupling technique is employed to separate generic and class-specific features. In addition, a binary gating mechanism is designed to dynamically assign experts while preventing over-reliance on specific categories. Extensive experiments on three benchmark long-tailed classification datasets:CIFAR10-LT, CIFAR100-LT, and ImageNet-LT-demonstrate that our method consistently outperforms existing state-of-the-art approaches. Ablation studies further confirm the effectiveness of the uniform enhanced sampling strategy and the joint optimization of multi-expert learning, showing that our algorithm successfully balances the model's attention across head and tail classes, thereby improving overall classification performance.

Graph convolution network is a powerful method of deep learning of graph structure data. Existing methods usually adjust the neighborhood information aggregation mode or optimize the graph topology layer by layer for improving the graph convolution network. However, these methods seldom consider the discriminative information about hierarchical characteristics nodes (some special nodes only can be correctly classified in one layer and are the misclassification nodes in the other layers of deep graph convolutional networks) in the different layers for complementing the neighborhood topology information. To further find these information, a deep graph layer information mining convolutional network (GLIM) can alternately measure the neighborhood ranking information on topology structure and update the residual identity mapping node information on the different layers for enhancing the model classification performance. Moreover, GLIM can construct a unified framework with the various hyper-parameters for the different graph learning method based on graph convolution network. Experiments show GLIM outperforms the state-of-the-art methods for semi-supervised node classification in three cite datasets (Cora, CiteSeer,and PubMed) and three image datasets (MNIST, Cifar10 and Cifar100).

How to propose an image retrieval algorithm with adaptable model and wide range of applications for large-scale datasets has become a critical technical problem in current image retrieval. This paper proposed an Incremental Image Retrieval Method Based on Feature Perception and Deep Hashing. The algorithm contains two important parts: the hash function learning part and the incremental hash code mapping part. Firstly, a module is designed called Feature Perception Module to obtain multi-scale global context-aware information. It also keeps the scale and shape of the final extracted deep features invariant. Then, a new incremental hash loss function is designed to maintain the similarity between the query image and the dataset image; the advantage of this is that it can reduce the time cost of updating the model. The experimental results show that the algorithm model can perform well in incremental image retrieval. It is shown that the algorithm can solve the current problem of low retrieval efficiency and high cost due to retraining models caused by the dramatic increase in the number of images in the image retrieval field.

Few-shot text style transfer faces two main challenges: The first challenge is the limited availability of reference style text, while the second challenge is the varying degrees of differences between the style reference image and the source image. Existing methods mainly focus on the influence of local style and global style feature extraction on text style, but they ignore the crucial role played by the difference between the style reference image and the source image on style characteristics, especially in Chinese, which has its own unique ideograph structure. To address this issue, this paper proposes Layer Similarity Guiding Few-shot Chinese Style Transfer (LSG-FCST). LSG-FCST can not only build a transfer network by encoding the content and style characteristics from low-level to high-level semantics, but it can also discover the similarity characteristics between the style reference image and the source image through the attention mechanism. Furthermore, LSG-FCST can integrate the style features generated by the similarity features of different layers and generate the target image through asymmetric decoding. In the self-built text image dataset, we consider three types of visibility situations for the test images in the training set: seen fonts unseen characters, unseen fonts seen characters, and unseen fonts unseen characters. The experiments show that LSG-FCST outperforms the state-of-the-art methods. The code and dataset can be accessed at https://github.com/LYM1111/LSG-FCST .

The fine-grained image retrieval task is challenging due to small inter-class and large intra-class variations, making it hard for hashing methods to distinguish subtle differences. Traditional algorithms often overlook detail in feature extraction, mapping similar images to different hash codes, while existing methods focus on dense regions and ignore other discriminative areas, lowering accuracy. Additionally, current approaches lack inter-channel correlation modeling, limiting feature aggregation. This paper proposes a hashing method based on salient region localization to capture fine-grained local features and an interactive channel transformation module to model inter-channel relationships. Experiments on four benchmark datasets show significant retrieval accuracy improvements over baselines, enhancing fine-grained image retrieval performance.

Aiming at the problems of color distortion, low image processing efficiency, rich context information, spatial information imbalance in the current low-light image enhancement algorithm based on a convolutional neural network. In this paper, an Attention-based multi-scale recursive residual network for low-light image enhancement (AMR-Net) is proposed based on high-resolution, single-scale image processing. First, shallow features are extracted using convolution and channel attention. In the recursive residual unit, a parallel multi-scale residual block is constructed, and the image features are extracted from the three scales: original image resolution, 1/2 resolution, and 1/4 resolution. Then, the deep features and shallow features are connected by selective kernel feature fusion to obtain rich context information and spatial information. Finally, the residual image is obtained by convolution processing of the deep features, and the enhanced image is obtained by adding the original image to the residual image. The experimental results on LOL, LIME, DICM, MEF datasets show that the proposed method has achieved good results in multiple indicators, and reasonably restored the brightness, contrast, and details of the image, thereby intuitively improving the perceived quality of the image.

Small target detection is an important research direction in the field of computer vision, which is widely used in popular fields such as industrial defect detection, satellite remote sensing image detection. However, in printing defects detection, due to the complex defect background and small target, it is difficult to extract multi-scale features, and the extracted features have less available information. Therefore, this paper improves YOLOv7 and proposes an end-to-end printing defects detection algorithm based on context structure perception and multi-scale feature fusion (CM-YOLOv7). CM-YOLOv7 is mainly composed of a Context Structure Awareness Module (CSAM), a Multi-scale Feature Interaction Module (MFIM) and a Feature Refinement Layer Module (FRLM). Firstly, the CSAM uses multiple convolution kernels of different sizes to obtain features of different receptive fields and enhance multi-scale feature extraction. Secondly, for the extracted multi-scale features, the MFIM adaptively fuses the features of adjacent layers to achieve mutual learning between coarse-grained information and fine-grained information and improves the expression ability of small target features lost after convolution. Finally, in order to refine the edge information of the defect target more effectively and enhance the feature expression ability of the image, an FRLM is designed in MFIM. In the experimental part, this paper utilizes the printing defects detection dataset and DOTA-V1.0 dataset in order to better evaluate the effect of the algorithm. The experimental results show that the CM-YOLOv7 model proposed in this paper has a more accurate detection effect on the small target of printing defects.

Occluded person re-identification aims to match occluded pedestrian with pedestrian images from nonintersecting cameras. For dealing with the task of Occluded Person Re-ID, existing methods focus only on tackling the pedestrian occlusion issue. Yet they all ignore the fact that salient nonhuman body parts such as hats, bags, and jewelry can provide more discriminatory informational cues. In this regard, this paper proposed an Occluded Person Re-ID algorithm based on Dual Attention Mask (DAM) guidance, which simultaneously solves the problems of pedestrian occlusion and significant non-human body part features that are easily ignored. Specifically, DAM consists of Pose Attention (PA) and Saliency Attention (SA). PA utilizes human pose key points to generate heatmap, and the produced heatmap indicates whether a specific body part is occluded or not, as well as guides the model to focus on the non-occluded region. SA then locates dropped salient non-human body part features through spatial attention and channel attention, and generates a saliency attention heat map to direct the model focus on the salient non-human body part features. Accordingly, DAM not only highlights the visible body parts while suppressing the occluded parts, but also localizes the prominent non-human body parts in the background. Finally, extensive experiments on three occlusion datasets demonstrate the effectiveness of our proposed method.