The Visual Place Recognition (VPR) task requires learning global representations that are both robust to severe appearance changes and discriminative enough to distinguish visually similar places. Most existing methods rely on a single global descriptor with fixed representation and aggregation strategies, which makes it difficult for one model to perform optimally across all appearance conditions commonly encountered in real-world environments. To address this limitation, we propose a Mixture-of-Experts (MoE) based global descriptor that enables adaptive representation learning for VPR. Our model consists of multiple aggregation branches and a lightweight routing network that performs soft adaptive weighting over all aggregation experts conditioned on the input image. This design allows the global descriptor to adapt to different visual characteristics while preserving the efficiency of one-stage retrieval. Extensive experiments on standard VPR benchmarks demonstrate that the proposed approach consistently outperforms strong baselines, with particularly notable improvements under challenging conditions such as day–night and seasonal changes.
Imperceptible adversarial attacks against face recognition systems require both effective identity deception and high visual fidelity. Recent diffusion-based attacks provide an effective framework for generating such adversarial samples. However, they still exhibit certain limitations, primarily characterized by indiscriminate perturbation allocation and insufficient consideration of structural preservation. In this paper, we propose a diffusion-guided partition reconstruction with structural invariance attack method (DPS-attack). To address the issue of indiscriminate perturbation allocation, we construct a fine-grained semantic prior derived from facial parsing and incorporate it as a region-aware optimization constraint. This prior regulates perturbation distribution across different facial regions, encouraging stronger preservation of sensitive areas while allowing more flexible optimization in less critical regions. To further balance attack effectiveness and structural fidelity, we design a dual-gradient guidance mechanism that integrates an identity-attacking gradient with a structure-preserving gradient during the denoising process. The former promotes identity deception, while the latter constrains geometric distortion by maintaining structural similarity. Extensive experimental results on the public FFHQ and CelebA-HQ datasets are compared with those of seven representative methods. Results demonstrate that DPS-attack achieves competitive attack success rates while improving the visual quality and imperceptibility of generated adversarial samples. Source code is available at https://github.com/zhanghrswpu/DPS-attack .
Cross-domain image retrieval (CDIR) under low-resource conditions is challenging because limited aligned data and large visual discrepancies hinder effective representation learning. This paper proposes a domain-gap embedding guided generative data augmentation framework for low-resource CDIR. The proposed method estimates the latent discrepancy between source and target domains and uses it to guide image synthesis, generating samples that preserve source-domain structure while incorporating target-domain visual characteristics. These synthetic samples enrich the training distribution and improve cross-domain feature alignment. Experiments on historical-map-to-satellite retrieval show that the proposed method improves early-rank retrieval performance over conventional augmentation and source-only generative augmentation under multiple low-resource settings. Additional evaluation on Sketchy sketch-to-photo retrieval further supports the applicability of the proposed framework to another heterogeneous domain pair. Ablation studies verify the importance of the target-guided domain-gap direction and the robustness of the proposed design. These results demonstrate that domain-gap guided generative augmentation is an effective data-centric strategy for low-resource CDIR.
The growing interest in image dehazing has led to significant advancements in deep learning-based single image dehazing approaches, achieving remarkable performance in recent studies. However, many existing methods fail to meet the efficiency demands of practical applications. In this paper, we introduce WaveDH, a novel and compact ConvNet designed to address this efficiency gap. Our WaveDH leverages wavelet sub-bands for guided up-and-downsampling and frequency-aware feature refinement. The key idea lies in utilizing wavelet decomposition to extract low- and high-frequency components from feature levels, allowing for faster processing while upholding high-quality reconstruction. The downsampling block employs a novel squeeze-and-attention scheme to optimize the feature downsampling process in a structurally compact manner, preserving discriminative features while discarding noise components. In our upsampling block, we introduce a dual-upsample and fusion mechanism to enhance high-frequency component awareness, aiding in the reconstruction of fine details. Departing from conventional dehazing methods that treat frequency components equally, our feature refinement block adopts a frequency-aware, coarse-to-fine approach to process features more effectively. By refining details in a maximum 8 × downsampled feature space, WaveDH strikes a favorable trade-off between accuracy and computational cost. Extensive experiments demonstrate that WaveDH outperforms many state-of-the-art methods across several benchmarks. Notably, on the SOTS-indoor dataset, WaveDH achieves a peak signal-to-noise ratio (PSNR) of 39.35 dB and a structural similarity index (SSIM) of 0.995, surpassing existing high-complexity models while requiring significantly reduced computational costs. Our code is available at https://github.com/smin-hwang/WaveDH .
This study introduces a computationally efficient and robust pipeline for rail defect detection that strategically combines shallow feature extraction with selective deep segmentation to achieve high precision with minimal computational overhead. We curated a dataset of 356 cropped rail surface images (132 Normal, 101 Joint, 123 Defective) from multiple public and private sources under varying real-world conditions, permanently partitioned into 249 training (70
