Weijian Hu

Accurate segmentation of retinal vessels is a key task in fundus image analysis and is crucial for the early screening and diagnosis of various ophthalmic and systemic diseases. However, existing methods still struggle to simultaneously address scale variation, contextual fusion, and the recovery of directional structures when dealing with retinal vessels, which easily leads to missed detection of thin vessels and structural discontinuities. To tackle these issues, we propose a retinal vessel segmentation network, MAF-Net. First, we design an adaptive multi-scale dilated residual block in the encoder, which performs branch-wise adaptive selection and reweighting over parallel dilated branches to dynamically adjust scale responses, thereby enabling adaptive modeling of vessel features with different calibers. Second, to enhance the network’s ability to understand context, we introduce a windowed hierarchical cross-scale attention fusion module at the bottleneck layer, which enables cross-scale interaction of multi-level features within local windows, thereby efficiently aggregating contextual information while reducing computational complexity. Finally, in the skip connections, we propose a semantically guided tri-axis fusion module, which extracts directional information along the horizontal, vertical, and spatial axes, and uses progressively guided features from lower-resolution layers to generate a spatial confidence map for gated selection, thereby suppressing shallow noise and reducing discontinuities and adhesion. On three public datasets, DRIVE, STARE, and CHASE-DB1, MAF-Net achieves accuracies of 97.14%, 97.99%, and 97.82%, respectively. These results indicate that MAF-Net improves vessel structure recovery while better preserving vascular connectivity and structural consistency.

Semantic segmentation of 3D point clouds remains difficult when LiDAR or depth-camera data are sampled unevenly. This paper presents DANet, a 3D semantic segmentation framework built on joint density- and semantics-adaptive convolution. Its core operator, Density-Adaptive Radius Convolution (DAR-Conv), predicts point-wise neighborhood radii before feature aggregation by combining density-driven initialization with semantics-aware modulation. In this way, dense regions can use compact receptive fields, whereas sparse or semantically complex regions can draw on broader contextual support. DANet also includes a Gated Adaptive Cross-Layer Fusion (GACF) module, which aligns encoder–decoder features and performs gated fusion with residual refinement. Experiments on S3DIS and NPM3D show that DANet obtains the highest reported mean accuracy (mAcc) among the compared methods on S3DIS, and high mean Intersection over Union (mIoU) and overall accuracy (OA) on NPM3D, supporting the usefulness of density- and semantics-aware receptive-field adaptation.

Accurate pixel-level segmentation of pavement cracks is an essential measurement task for road safety and infrastructure maintenance, as well as for the quantitative measurement of crack morphology that supports condition assessment and maintenance planning. Although deep learning has advanced crack segmentation, existing methods still struggle to capture long, continuous cracks and delineate irregular edges under complex backgrounds, which limits measurement reliability. To address these challenges, we propose HiResCrackNet, a high-resolution segmentation model designed to enhance both segmentation accuracy and measurement robustness. The model employs a high-resolution backbone to preserve fine-grained features, a hierarchical feature adjuster to balance global structure and local detail, and a multi-scale fusion perception module to strengthen cross-level feature integration. In addition, we release BT-Crack2000, a large-scale dataset with diverse and challenging scenarios to support robust model training and benchmarking. Extensive experiments across multiple datasets demonstrate that HiResCrackNet consistently outperforms state-of-the-art methods, delivering precise crack delineation. These contributions provide a reliable foundation for quantitative pavement evaluation and highlight the potential of intelligent measurement approaches in smart infrastructure management.

Neural machine translation (NMT) systems have achieved outstanding performance and have been widely deployed in the real world. However, the undertranslation problem caused by the distribution of high-translation-entropy words in source sentences still exists, and can be aggravated by poisoning attacks. In this paper, we propose a new backdoor attack on NMT models by poisoning a small fraction of parallel training data. Our attack increases the translation entropy of words after injecting a backdoor trigger, making them more easily discarded by NMT. The final translation is part of target translation, and the position of the injected trigger poison affects the scope of the truncation. Moreover, we also propose a defense method, Backdoor Defense by Sematic Representation Change (BDSRC), against our attack. Specifically, we selected backdoor candidates based on the similarity between the semantic representation of words in a sentence and the overall sentence representation. Then, the injected backdoor is identified through computing the semantic deviation caused by backdoor candidates. The experiments show that our attack strategy can achieve a nearly 100% attack success rate, and the functionality of main translation tasks is almost unaffected in models having performance degradation that is less than 1 BLEU. Nonetheless, our defense method can effectively identify backdoor triggers and alleviate performance degradation.