Accurate segmentation of colorectal polyps can effectively prevent the incidence of colorectal cancer. Current segmentation methods primarily rely on spatial domain analysis with insufficient frequency information utilization. This spatial-centric approach suffers from locality constraints and limited sensitivity to subtle polyp variations. Furthermore, existing techniques often exhibit inadequate information transmission and fusion across multi-scale features. To address these issues, we propose DCSFNet, a colon polyp segmentation method based on joint spatial-frequency domain perception and deeply coupled feature interaction, which includes four key components: multi-scale feature reconstruction optimization module (MROM), dual-domain group learning module (DGLM), dual-domain interactive perception module (DIPM) and multi-scale feature global perception aggregation module (MGPAM). Specifically, MROM is introduced to reconstruct encoder features and progressively transmit the local fine details of polyps, mitigating information loss in high-level semantics. DGLM is designed to simultaneously enhance local structural details and global contextual understanding by exploiting complementary cues from both spatial and frequency domains. Furthermore, DIPM is present to facilitate effective interaction and complementary learning between spatial and frequency domain features, providing richer and more comprehensive information for segmentation. Finally, MGPAM is proposed to aggregate multi-scale features to capture richer contextual interdependencies. Comprehensive experiments demonstrate that DCSFNet achieves superior segmentation performance across five benchmark polyp datasets.

Existing chaos-based color image encryption algorithms still face several challenges, including insufficient dynamical complexity of low-dimensional chaotic maps, residual boundary regularity caused by fixed block partitioning, and limited diffusion among RGB channels. To address these issues, this paper proposes a color image encryption algorithm based on a three-dimensional sine-bilinear fully coupled chaotic map (3D-SBFCM). The proposed map integrates sinusoidal modulation, linear coupling, and bilinear cross-coupling within a mod-1 mapping framework, thereby improving the complexity and pseudorandomness of the generated chaotic sequences. In addition, a residual-feasibility-constrained dynamic rectangular partitioning mechanism is developed to generate reversible non-uniform image blocks and reduce the structural regularity associated with fixed-size partitioning. Based on this partitioning structure, inter-block permutation among same-size blocks and intra-block two-dimensional permutation are performed to weaken both global and local spatial correlations. Plaintext-related initialization, dynamic S-box substitution, and forward-backward cross-channel diffusion are further incorporated into the overall permutation-diffusion framework to enhance plaintext sensitivity, nonlinear confusion, and perturbation propagation across RGB channels. Experimental results demonstrate that the proposed algorithm effectively conceals the statistical characteristics of plaintext images, with information entropy values higher than 7.999 for all color channels and NPCR/UACI values close to their theoretical expectations. The algorithm also shows satisfactory robustness against cropping and noise attacks. These results indicate that the proposed method provides an effective and secure solution for color image encryption.

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 https://github.com/Pengxiaolong293/FDESNet.

In order to solve the problem of low security in image encryption algorithms based on a single chaotic map, this paper proposes a new chaotic system, namely, a cascaded modulated sine chaotic system (CMSCS) that can generate several different chaotic maps simultaneously. Based on this, we propose a block-based image encryption algorithm that leverages the characteristics of CMSCS to achieve one encryption at a time. Image encryption involves three stages. First, the CMSCS is constructed by introducing existing chaotic maps as seed maps, cascading them with nonlinear functions, and then performing a sine transform on them. The SHA-512 hash function is used to generate a key associated with the plaintext, which is then employed to produce a chaotic sequence with strong randomness. Next, a spiral scrambling algorithm is devised, which achieves pixel position swapping by rotating pixels clockwise and is combined with a chaotic-sequence-based scrambling algorithm to implement double scrambling of the plaintext. Finally, by integrating the two-dimensional index matrix position transformation method, we design a snake-shaped transform and a Latin square-like transform to achieve multiple nonlinear diffusion operations both within and between blocks. Simulation results show that the proposed encryption algorithm can effectively resist typical attacks such as brute-force attack, differential analysis, and statistical analysis, and exhibits high security.

While deep learning has achieved significant progress in medical image segmentation, its reliance on large-scale annotated data remains a major bottleneck in limited-sample scenarios. Few-shot medical image segmentation (FSMIS) aims to address this challenge by enabling effective segmentation under data-scarce conditions. However, existing prototype-based methods typically generate prototypes from support features via random sampling or local averaging, thereby overlooking the query-specific demands. To this end, we propose ReLiFSS (Reliability-Aware Few-Shot Medical Image Segmentation), a query-feature-guided method. Its core idea is to directly integrate query features into the prototype generation process to construct customized prototypes tailored to different query images. Specifically, we design a Query Pseudo-Prototype Generation (QPPG) module, which constructs an initial prototype using support features and performs preliminary segmentation on the query image to extract a pseudo-prototype reflecting its specific requirements. Subsequently, the High-Confidence Support Prototype Generation (HSPG) module utilizes this pseudo-prototype to reverse-segment the support set, mining the feature regions discriminative for the query image segmentation. Finally, the Multiple Prototypes Matching Segmentation (MPMS) module fuses foreground and background information via a dual-pathway mechanism, helping alleviate foreground-background imbalance. Extensive experiments on three public medical image datasets show that our method achieves competitive performance compared with recent FSMIS methods. The code is available at: https://github.com/caihanyue49-art/ReLiFSS .

The malicious dissemination of fake images has triggered a societal trust crisis, making deepfake detection a critical area of research. However, existing methods often rely on single-domain features, which limits their generalization to diverse forgery techniques. To address this limitation, this paper proposes a novel Multi-Feature Contrastive Learning (MFCL) model for deepfake detection. MFCL integrates multiple image features within an unsupervised contrastive learning framework. Specifically, 15 feature combinations are designed based on six image features from three domains: color (RGB, YCbCr), texture (Sobel, Prewitt), and frequency (DCT, SRM), aiming to identify the most effective multi-feature configuration for MFCL. Furthermore, a Feature Mix Module (FMix) is introduced to enhance feature fusion through depthwise separable convolutions. Extensive experiments on multiple datasets demonstrate that MFCL achieves superior detection accuracy and robustness, particularly in cross-dataset and cross-manipulation scenarios, highlighting its strong generalization capabilities.

With the increasing popularity of LiDAR in autonomous driving, 3D object detection based on point cloud has become a research hotspot in both industry and academia. However, most methods are limited by the inherent challenges of point cloud representation and sparsity, which makes it difficult to balance computational efficiency and accuracy. To address these challenges, this paper proposes a multimodal 3D object detection method, termed PV-MM3D. Specifically, we design a Point-Voxel Parallel Dual-Stream Framework, which leverages independent point- and voxel-based dual-stream networks to process virtual and LiDAR point clouds in parallel. This design preserves the strengths of point-based methods in capturing the intricate 3D structures of objects while exploiting the computational speed advantage of voxel-based methods. To improve the sampling proportion of foreground points, we introduce a Dynamic Dual-Sampling Mechanism, which dynamically downsamples the point cloud in a learnable manner to reduce background points while retaining key foreground information. To maintain the independence of features while achieving cross-modal feature interaction, we present a Dual-Attention Region Adaptive Fusion Module, which utilizes the attention mechanism to guide the network in adaptively adjusting the importance between point-voxel multimodal features. Experiments on the KITTI dataset demonstrate that PV-MM3D achieves mAP of 83.99 % for 3D detection and 91.31 % for BEV detection in the car category. The code is available at https://github.com/BaotWang/PV-MM3D.

Accurate polyp segmentation from colonoscopy images is pivotal for the early detection of colorectal cancer (CRC), significantly enhancing diagnostic efficiency and reliability in clinical practice. While recent methods have achieved notable progress, they often suffer from two critical limitations: (1) inadequate frequency and spatial feature representation, as most approaches remain biased toward spatial-domain learning and, even when incorporating frequency information, tend to overlook the hierarchical variability of frequency distributions across feature levels, resulting in suboptimal utilization of frequency cues; and (2) insufficient cross-level feature integration, limiting the ability to effectively capture both global semantics and fine-grained boundary details. To address these issues, we propose PSFS-Net, a novel dynamic frequency-spatial synergistic polyp segmentation framework that jointly leverages spatial and frequency domain information for hierarchical context refinement and cross-level fusion, which mainly includes Frequency-aware Cross-scale Fusion Module (FACFM), Dual-stream Global-Local Interaction Module (DGIM), and Dual Attention Cross-modulation Module (DCM). Specifically, FACFM is designed to extract frequency domain cues and adaptively decoupling high/lowfrequency components from full-spectrum information by employs Discrete Fourier Transform and an adaptive Dynamic Gaussian Filters. DGIM is introduced to enable mutual refinement between high-level semantic representations and low-level spatial details through dedicated global and local processing branches. DCM is presented to further aggregate global contexts with local details via dual-attention mechanisms, alleviating semantic gaps. Extensive evaluations on five public polyp segmentation datasets demonstrate that PSFS-Net delivers competitive and excellent performances. Our code is available at https://github.com/chljzh25/PSFS-Net.

Recommender systems play a vital role in enhancing user experience by efficiently delivering personalized and relevant content. While knowledge graph-based recommender systems effectively alleviate the data sparsity and cold-start challenges of traditional approaches, they still suffer from two major limitations: (1) insufficient utilization of the user-item interaction matrix and (2) suboptimal integration of heterogeneous knowledge graph signals with collaborative information. In this work, we propose DAIGNN (Dual-Adaptive Imputation Graph Neural Network), a novel recommendation framework designed to overcome these limitations through three key innovations. First, we introduce a similarity-driven imputation mechanism that constructs an Imputation Graph using pseudo-ratings, thereby enhancing graph connectivity and significantly reducing data sparsity. Second, we incorporate multiple auxiliary information sources on both the user and item sides, enabling DAIGNN to capture richer contextual and relational semantics beyond conventional user-item interactions. Third, we develop a dual-adaptive feature fusion mechanism that learns optimal fusion weights to dynamically integrate heterogeneous information from multiple graph sources. Extensive experiments conducted on four real-world datasets demonstrate the superior effectiveness of DAIGNN. On average, it achieves a 3.1% improvement in AUC and a 2.0% improvement in F1-score over state-of-the-art baselines, confirming its robustness across diverse settings.

Arbitrary style transfer has become a prominent research focus in the field of computer vision. However, the strong coupling between image style and content structural information poses significant challenges for effective feature extraction when using a single VGG network. In addition, the low semantic correlation between content and style often leads to distorted structures or over-stylization in the generated results. To address these issues and produce high-quality stylized images, we propose a Semantic Sparse and Dense Adaptive Fusion Network (SSDAF-Net). Specifically, we design a Grouped Self-Attention (GSA) module that adaptively extracts localized style representations along the channel dimension within different groups of the style image. By integrating channel grouping with dilated convolution, the GSA effectively suppresses excessive inter-channel interference commonly observed in global self-attention, thereby enriching the intrinsic style representation. Furthermore, to handle the low semantic relevance between content and style images that may result in unrealistic stylization patterns, we develop an Adaptive Fusion Attention Module (AFAM), which effectively processes semantically weakly correlated regions through sparse and dense attention branches. Prior to fusion, an Intrinsic Content Purification Module (ICPM) is introduced to deeply refine and disentangle the intrinsic semantic features of content and style images, providing precise semantic information for subsequent fusion. Extensive qualitative and quantitative experiments demonstrate that the proposed method generates high-quality stylized images and achieves superior performance in arbitrary style transfer tasks.

Self-supervised monocular depth estimation that does not require ground truth for training has attracted attention in recent years. However, existing approaches still face challenges such as inaccurate predictions and blurred edges when handling regions with abrupt structural changes, including object boundaries and occluded areas. To address these problems, we propose a structure-aware guided self-supervised monocular depth estimation framework, focusing on improving depth prediction accuracy in edge regions. This method comprises three key modules: the Edge-Aware Feature Optimizer (EAFO), which predicts edge probability maps from high-level semantics of the encoder and guides feature enhancement in intermediate layers to strengthen boundary representations; the Adaptive Feature Enhancement Module (AFEM), which aggregates local and global semantic information using a dilated attention mechanism to maintain consistency in edge regions; and the Wavelet Domain Decoder Block (WBD), which introduces frequency-domain modeling during the decoding phase, achieving clear and structurally complete depth recovery through fusion of low-and high-frequency structural differences and attention guidance.Experimental results demonstrate that the proposed method achieves superior performance on datasets such as KITTI and Make3D, particularly outperforming existing mainstream self-supervised methods in edge structure regions. Meanwhile, it maintains low computational cost and parameter count, exhibiting excellent generalization ability and deployment potential.

Medical image segmentation plays a pivotal role in clinical diagnosis and biomedical research. However, Convolutional Neural Networks (CNNs) struggle to capture long-range dependencies, whereas Transformers—despite their global modeling capability—incur quadratic computational complexity. Recently, State-Space Models (SSMs), exemplified by Mamba, have enabled linear-complexity modeling of long-range dependencies. However, existing Mamba-based models still exhibit limitations in capturing local textures and edges and modeling spatial continuity. To address this challenge, we propose LEAP-Mamba (Locality-Enhanced Feature Calibration with Pixel–Region Dual-Stream Vision Mamba UNet). Specifically, we introduce a Pixel–Region VSSblock that comprises a local-Mamba branch and a global-Mamba branch to simultaneously capture fine-grained details and high-level semantic context. Furthermore, we design a Hierarchical Feature Calibration Module (HFCM) that leverages multi-scale context awareness and a structural saliency filtering mechanism to dynamically calibrate the skip connections, effectively suppressing feature redundancy while enhancing semantic interaction quality. Additionally, we propose a Cross-Region Feature Interaction (CRFI) module that aggregates spatial context via clustering to facilitate efficient inter-region interaction and semantic sharing. Extensive experiments on the ISIC2016, ISIC2017, ISIC2018, and Synapse datasets demonstrate that LEAP-Mamba achieves competitive performance compared with state-of-the-art methods, with leading results on several key metrics and notable improvements in challenging categories such as small organs and low-contrast structures. The code will be made available at https://github.com/caihanyue49-art/LEAP-Mamba .

With its wide applications in artistic creation and visual generation, image style transfer has become an important research topic in the field of computer vision. However, existing methods still suffer from issues such as pattern repetition and incomplete structural preservation during stylization. To address these problems, we propose a Cross-level Semantic Alignment and Structure Preservation Network (CASPNet). First, a Semantic Refinement and Synergistic Fusion Module (SRSFM) is designed, which leverages channel attention to refine style features and suppress redundant background noise. It further performs cross-level semantic alignment from both local and global perspectives, while introducing a structure-aware spatial gating mechanism to enhance the preservation of content structures. Second, to improve the style representation capability during the decoding stage, a Style Modulation Module (SMM) is constructed. It utilizes shallow style features to generate dynamic scaling and shifting parameters, which are applied to modulate the fused features during decoding, thereby effectively alleviating the loss of style information caused by upsampling. Finally, a local structure loss is designed to preserve more detailed content structures. Both qualitative and quantitative experimental results demonstrate that CASPNet achieves significant improvements in mitigating pattern repetition and preserving structural integrity, producing high-quality stylized images. The code is available at https://github.com/zkzhu02/CASPNet .

Deep hashing technology has gained widespread attention due to its low storage cost and high retrieval efficiency. Although existing cross-modal hashing methods have achieved good retrieval results, there is still the problem of facing modal heterogeneity leading to semantic separation. Therefore, this paper proposes a deep hashing method based on prototype-aware hardness-weighted, which uses the CLIP model after fine-tuning the large language model as a feature extractor to better obtain the feature information of the two modalities. And a set of loss functions are designed to better handle difficult samples by prototype-aware hardness-weighted loss to ensure that the data can be embedded in the appropriate location, bringing relevant data closer and pulling irrelevant data away. After conducting experiments on three datasets and comparing with some advanced cross-modal retrieval methods in recent years, it can be shown that our proposed DPHS method has excellent performance. The code and datasets used in this article can be obtained from https://github.com/chmy180/dphs-main .

Monocular depth estimation (MDE) seeks to infer pixel-wise dense depth maps from a single RGB image. Recent methodologies predominantly utilize the encoder-decoder architecture to effectively extract and analyze multi-scale features. However, they tend to ignore the important role that high-level features with rich global information play in MDE, resulting in a poor understanding of the overall structure of the scene by the model. Based on this, we propose a novel encoder-decoder framework called GRdepth, which includes a cross large scale feature enhancement (CLSE) module and an iterative regulation decoder (IRD). Specifically, the CLSE module is designed to use high-level features, enriched with global information extracted by a global information aggregation (GIA) unit, to guide the enhancement of multi-scale feature maps produced by the encoder. This enhancement is achieved through a cross large scale feature fusion (CLSF) unit built from channel attention and spatial attention to refine low-level features with high-level information. The IRD is tailored for MDE based on classification-regression which mainly utilizes a bin width self-regulation (SRbins) unit to adjust the width of the initial bins predicted with the bottleneck features. This adjustment is guided by bin width predicted by an iterative adaptive feature fusion (IAFF) unit at each level, effectively combining global information and local information for more accurate bin width and bin centers. Extensive experiments on the indoor dataset NYU-Depth-v2 and SUN-RGBD and on the outdoor dataset KITTI demonstrate that our method can achieve comparable state-of-the-art (SOTA) results.

Clothes-Changing Person Re-identification (CCReID) aims to match identities across images of individuals in different attires. Due to the significant appearance variations caused by clothing changes, distinguishing the same identity becomes challenging, while the differences between distinct individuals are often subtle. To address this, we reduce the impact of clothing information on identity judgment by introducing linguistic modalities. Considering CLIP’s (Contrastive Language-Image Pre-training) ability to align high-level semantic information with visual features, we propose a CLIP-driven Semantic Guidance Network (CSGN), which consists of a Multi-Description Generator (MDG), a Visual Semantic Steering module (VSS), and a Heterogeneous Semantic Fusion loss (HSF). Specifically, to mitigate the color sensitivity of CLIP’s text encoder, we design the MDG to generate pseudo-text in both RGB and grayscale modalities, incorporating a combined loss function for text-image mutuality. This helps reduce the encoder’s bias towards color. Additionally, to improve the CLIP visual encoder’s ability to extract identity-independent features, we construct the VSS, which combines ResNet and ViT feature extractors to enhance visual feature extraction. Finally, recognizing the complementary nature of semantics in heterogeneous descriptions, we use HSF, which constrains visual features by focusing not only on pseudo-text derived from RGB but also on pseudo-text derived from grayscale, thereby mitigating the influence of clothing information. Experimental results show that our method outperforms existing state-of-the-art approaches.

In point cloud 3D object detection tasks, single-stage detectors offer fast inference but are less accurate than two-stage detectors. We point out two main problems: first, traditional methods deal with the whole point cloud, making them vulnerable to background noise interference; second, existing methods exhibit insufficient single-channel feature encoding capability. Therefore, this paper proposes Instance-Aware Sampling and VoxelTransformer Encoding for Single-Stage 3D Object Detection (IAVT-SSD). Specifically, we design an Instance- Aware Weighted Sampling Strategy to filter out ground reflection points, enhancing the model's focus on the foreground points. Meanwhile, we introduce a Voxel-Transformer Dual-Channel Feature Encoding Module to capture more comprehensive features through two independent channels, efficiently fusing non-empty voxels and remote context information. In addition, a Collaborative Enhancement Branch is designed to predict the complete structure of the object. Experiments show that IAVT-SSD achieves a good balance of accuracy and speed, with an inference speed of 42 FPS (frames per second) and a mAP (mean average precision) of 81.70% on the KITTI dataset, and a mAP of 66.96% on the ONCE dataset, validating its effectiveness and superiority.

Facial Expression Recognition (FER) with occlusion presents a challenging task in computer vision because facial occlusions result in poor visual data features. Recently, the region attention technique has been introduced to address this problem by researchers, which make the model perceive occluded regions of the face and prioritize the most discriminative non-occluded regions. However, in real-world scenarios, facial images are influenced by various factors, including hair, masks and sunglasses, making it difficult to extract high-quality features from these occluded facial images. This inevitably limits the effectiveness of attention mechanisms. In this paper, we observe a correlation in facial emotion features from the same image, both with and without occlusion. This correlation contributes to addressing the issue of facial occlusions. To this end, we propose a Information Gap based Knowledge Distillation (IGKD) to explore the latent relationship. Specifically, our approach involves feeding non-occluded and masked images into separate teacher and student networks. Due to the incomplete emotion information in the masked images, there exists an information gap between the teacher and student networks. During training, we aim to minimize this gap to enable the student network to learn this relationship. To enhance the teacher's guidance, we introduce a joint learning strategy where the teacher conducts knowledge distillation on the student during the training of the teacher. Additionally, we introduce two novel constraints, called knowledge learn and knowledge feedback loss, to supervise and optimize both the teacher and student networks. The reported experimental results show that IGKD outperforms other algorithms on four benchmark datasets. Specifically, our IGKD achieves 87.57% on Occlusion-RAF-DB, 87.33% on Occlusion-FERPlus, 64.86% on Occlusion-AffectNet, and 73.25% on FED-RO, clearly demonstrating its effectiveness and robustness. Source code is released at: https://github.com/Lzh-Code-lab/IGKD.