Industrial anomaly detection is a critical task in automated visual inspection systems, aiming to identify subtle defects without relying on large-scale anomaly annotations. Recently, pretrained vision transformers such as DINO have demonstrated strong feature representation capabilities and have been widely adopted in this domain. However, most existing methods rely on single-grid patch partitioning, which suffers from boundary information loss and limited local context modeling, resulting in degraded performance on small or boundary-located defects. In this paper, we propose a simple yet effective framework built on a frozen DINOv3 ViT-B/16 backbone for industrial anomaly detection. To address the limitations of conventional patch-based inference, we introduce a Multi-view Patch Inference (MVPI) strategy, which constructs multiple overlapping spatial views of an input image. Each view is encoded independently by DINOv3, and overlapping regions are merged by a deterministic mid-point rule, so that no output token lies on a view boundary. On top of the extracted features, we further propose a Multi-scale Reconstruction Fusion Transformer (MRFT), in which lightweight reconstruction heads model the normal feature distribution of each backbone layer via cosine similarity-based reconstruction learning. The anomaly score is computed from reconstruction residuals, enabling effective discrimination between normal and defective regions. Extensive experiments on three benchmark datasets, including MVTec AD, BTAD, and UltraChip, show that the proposed method achieves state-of-the-art or competitive performance in both image-level and pixel-level anomaly detection, while requiring no memory bank and only moderate GPU memory at inference. The code will be released at https://github.com/chaoqunWang666 .
Images captured under poor lighting conditions typically suffer from low visibility and inherent sensor noise. Despite the notable progress of existing self-supervised methods, jointly achieving reliable illumination enhancement, effective noise suppression, and strong generalization remains challenging. To this end, we propose a novel self-supervised low-light image enhancement (LIE) method with neighbor sampling and mix-up consistency, termed NSMLIE, built upon Retinex decomposition. To improve brightness without paired supervision, we design a set of self-supervised constraints to guide stable Retinex decomposition. To further improve robustness against sensor noise and diverse illumination conditions, we propose a reflectance consistency constraint that integrates neighbor sampling with a cross-illumination mix-up mechanism. This design encourages consistent reflectance estimation across spatially related regions and different illumination conditions, thereby facilitating noise-robust reflectance recovery under varying lighting environments. Extensive experiments on three widely used LIE benchmark datasets demonstrate that the proposed NSMLIE effectively enhances low-light images while suppressing noise artifacts, delivering strong overall performance against 13 existing methods and yielding the best or second-best SSIM and LPIPS scores. Codes are available at https://github.com/yangyan22/NSMLIE.
The automated extraction of visual information from Dunhuang murals is of profound significance for the preservation and semantic analysis of digital cultural heritage. However, conventional models often suffer from performance degradation due to physical deterioration, chaotic background noise, and highly imbalanced data distributions. To address these challenges, this paper proposes a semantic-enhanced framework tailored for complex semantics and long-tailed categories. A central component of the framework is the Global-Local Semantic Enhancement Module (GL-SEM), designed to bridge the gap between pixel-level visual features and high-level compositional logic. Specifically, the Local Semantic Enhancement unit (L-SEM) utilizes a Simplified Convolutional Block Attention Mechanism (SCBAM) to adaptively refine features while suppressing semantic noise within cluttered backgrounds. Subsequently, the Global Semantic Reasoning unit (G-SEM) incorporates an Asymmetric Dense Relation Block (ADRB) to model cross-regional contextual dependencies, enabling the framework to perform contextual joint reasoning by establishing relational links between disparate visual entities. Furthermore, a deformation-adaptive feature reconstruction module is introduced to handle irregular geometric distortions, and a Dynamic Weighted Loss (DWL) strategy is employed to prioritize the extraction of low-frequency rare elements, thereby alleviating the long-tail effect. Experimental results demonstrate significant improvements over the baseline, with recall and mAP@0.5 increasing by 4.5 and 5.1
The 3D Gaussian Splatting (3DGS) approach provides an efficient solution for novel view synthesis through explicit scene representation and differentiable rasterization. While feed-forward approaches improve the efficiency of sparse-view reconstruction, their shared latent spaces often entangle geometry, appearance, and uncertainty representations, which may affect stable feature learning in occluded and textureless regions. In this paper, we present GAUSplat, a disentangled latent Gaussian representation that separates geometry, appearance, and uncertainty components to improve representation consistency under sparse and ambiguous observations. Based on the proposed representation, we develop a cross-view alignment strategy with uncertainty-aware feature regularization and a lightweight residual appearance modulation module. Experiments on RealEstate10K demonstrate that GAUSplat improves rendering consistency and perceptual quality under sparse-view settings, while achieving a favorable balance between reconstruction quality and computational efficiency.
Personalized recommendation systems have become essential in various applications, yet conventional collaborative filtering (CF)-based methods often struggle to capture deep semantic associations due to the absence of explicit knowledge guidance. Furthermore, current knowledge graph (KG)-based methods introduce redundant task-irrelevant knowledge, hindering interest learning and recommendation performance. To address these limitations, we propose a knowledge-guided interest co-teaching (KICT) framework for personalized recommendation, which aims to comprehensively and accurately model users’ interests. KICT consists of two key modules: dual-channel graph encoding and bidirectional co-teaching alignment. Dual-channel graph encoding jointly learns and propagates user interests from both collaborative and semantic perspectives, while bidirectional co-teaching alignment enables cross-view interest-aware contrastive supervision, achieving semantic enrichment while filtering task-irrelevant KG noise. Extensive experiments demonstrate that KICT significantly outperforms existing CF and KG-based baselines, highlighting the effectiveness of the proposed approach.
