Huiqin Wang

UAV-based bridge crack inspection requires efficient crack localization and accurate mask extraction under lightweight deployment constraints. This paper proposes a lightweight Direction-Aware C3k2 (DA-C3k2) network for bridge crack detection and instance segmentation. The proposed method redesigns the C3k2 feature extraction block into a Direction-Aware C3k2 block, which integrates multi-orientation feature decomposition, coordinate-guided spatial calibration, and gated residual fusion. The resulting shared backbone learns direction-aware crack representations, while the detection and segmentation heads are supervised by task-specific objectives for coarse crack localization and fine-grained mask prediction, respectively. Tversky loss is further introduced for sparse crack masks to reduce the foreground–background imbalance in thin crack segmentation. On the UAV-PDD2023 crack subset, the final DA-C3k2 detector improves mAP50 from 0.670 to 0.749 and mAP50–95 from 0.334 to 0.382 compared with the YOLO11n baseline, while under the same final training protocol DA-C3k2 improves the C3k2 baseline from 0.733 to 0.749 mAP50 and from 0.373 to 0.382 mAP50–95. For segmentation, DA-C3k2 with Tversky loss achieves the best mask mAP among the evaluated lightweight variants on the original benchmark, and obtains 0.5225 Dice, 0.3811 IoU, 0.5182 boundary-F1, and 0.3439 boundary-IoU on an external COCO-format crack dataset at 7.20 FPS under CPU inference. These results indicate that direction-aware feature modeling and foreground-sensitive supervision improve lightweight crack localization and mask extraction for UAV bridge inspection.

Scene graph generation (SGG) provides a structured representation for visual understanding. However, most existing methods are designed to optimize global triplet recall rather than retrieve relation instances specified by a user query. In query-driven visual relation retrieval, two major challenges arise: the target relation must compete with a highly redundant candidate space, and query semantics are not incorporated before relation classification. To address these challenges, we propose a Query-Driven Candidate Relation Screening (QCRS) module, which injects query semantics into the candidate screening process. Specifically, QCRS encodes the query and candidate visual relation features, and then filters query-relevant candidates through relevance scoring. By reducing interference from irrelevant candidates and avoiding redundant computation, QCRS improves the final exact triplet hit performance and enhances the interpretability of query-specific relations, thereby facilitating query-driven visual relation retrieval. Built upon the strong EGTR baseline, QCRS learns query relevance to prioritize relation instances matching the target query, enabling precise triplet retrieval. Extensive ablation studies and analyses on the VG150 benchmark validate the effectiveness of the proposed approach: when integrated with EGTR, QCRS improves PairR@50 from 61.52% to 80.06% and ETR@50 from 30.54% to 47.07%, achieving absolute gains of over 16 percentage points in both correct object pair retention and end-to-end target relation retrieval performance.

Among the approximately 60,000 bone stick fragments unearthed from the Weiyang Palace site of the Han Dynasty, about 57,000 bear inscriptions. Most of these fragments exhibit vertical fractures, leading to a separation between the upper and lower fragments, which poses significant challenges to digital preservation and artifact restoration. Manual matching is inefficient and may cause further damage to the bone sticks. This paper proposes a novel multimodal bone stick matching approach that integrates image, inscription, and archeological information to enhance the accuracy and efficiency of matching fragmented bone stick artifacts. Unlike traditional methods that rely solely on image data, our method leverages large-scale pre-trained models, namely Vision-RWKV for visual feature extraction, RWKV for inscription analysis, and BERT for archeological metadata encoding. A dynamic cross-modal feature fusion mechanism is introduced to effectively combine these features, enabling better interaction and weighting based on the contextual relevance of each modality. This approach significantly improves matching performance, particularly in challenging cases involving fractures, corrosion, and missing sections. The novelty of this method lies in its ability to simultaneously extract and fuse multiple sources of information, addressing the limitations of traditional image-based matching methods. This paper uses Rank-N and Cumulative Match Characteristic (CMC) curves as evaluation metrics. Experimental evaluation shows that the matching accuracy reaches 94.73% at Rank-15, and the method performs significantly better than the comparative methods on the CMC evaluation curve, demonstrating outstanding performance. Overall, this approach significantly enhances the efficiency and accuracy of bone stick artifact matching, providing robust technical support for the research and restoration of bone stick cultural heritage.

The stitching of bone stick fragments is of great significance for the inheritance of and research into the outstanding traditional culture of the Western Han Dynasty. Focused on the problem that the existing methods have a low bone stick fragment stitching success rate due to their complex edges, this paper proposes a bone stick fragment stitching method based on improved corner detection, which strengthens the features of broken edges and improves the success rate of stitching. First, the maximum outer contour of the bone stick fragment is obtained using connectivity analysis. Then, the broken edge features are enhanced by combining multi-scale and multi-stage corner detection, as well as a double screening mechanism of local template and corner density analysis. On this basis, the maximum outer contour is segmented using a region segmentation method based on corner features. Then, a priority matching strategy is used to perform multi-dimensional sequential priority matching of corner features for the densest segments and to match other contour segments in the order of density. Finally, based on the matching results, the bone stick fragment is stitched exactly and restored. Experimental results show that compared with other algorithms, the average proportion of broken corners when using the method in this paper is increased by about 44.89%, and the average comprehensive weighted similarity is increased by about 32.42%, which meets the needs of bone stick fragment stitching.

The unearthed artifacts from the ruins of the Weiyang Palace in the Han Dynasty’s Chang’an City include a large number of bone stick relics. The front and back of each bone stick form a pair, with similar appearances within each pair, indicating a matching relationship. However, the close resemblance in shape and complex texture features of the bone sticks make traditional image matching methods inefficient and less accurate. To address these challenges, this paper proposes a ConvNeXt-based Siamese network for bone stick image matching, utilizing feature similarity metrics. The method introduces a polarized self-attention module to enhance the model’s focus on the edges and texture features of the bone sticks, thereby improving the feature extraction capability of the ConvNeXt network. The utilization of DIS-Net for background normalization reduces noise interference, enhancing the precision of model matching. The adaptive adjustment of the learning rate using the cosine annealing algorithm accelerates network convergence. Additionally, transfer learning and contrastive learning methods are employed to pre-train the network, expediting convergence and improving matching accuracy. Experimental results demonstrate that compared to Siamese network algorithms with backbone networks such as VGG16, ResNet50, and MobileNet, the proposed method achieves an accuracy of 88.46% and 80% at Rank-10. This indicates a higher matching accuracy, providing an effective solution for automated bone stick matching.

The rejoining of bone sticks holds significant importance in studying the historical and cultural aspects of the Han Dynasty. Currently, the rejoining work of bone inscriptions heavily relies on manual efforts by experts, demanding a considerable amount of time and energy. This paper introduces a multi-scale feature fusion Siamese network guided by edge contour (MFS-GC) model. Constructing a Siamese network framework, it first uses a residual network to extract features of bone sticks, which is followed by computing the L2 distance for similarity measurement. During the extraction of feature vectors using the residual network, the BN layer tends to lose contour detail information, resulting in less conspicuous feature extraction, especially along fractured edges. To address this issue, the Spatially Adaptive DEnormalization (SPADE) model is employed to guide the normalization of contour images of bone sticks. This ensures that the network can learn multi-scale boundary contour features at each layer. Finally, the extracted multi-scale fused features undergo similarity measurement for local matching of bone stick fragment images. Additionally, a Conjugable Bone Stick Dataset (CBSD) is constructed. In the experimental validation phase, the MFS-GC algorithm is compared with classical similarity calculation methods in terms of precision, recall, and miss detection rate. The experiments demonstrate that the MFS-GC algorithm achieves an average accuracy of 95.5% in the Top-15 on the CBSD. The findings of this research can contribute to solving the rejoining issues of bone sticks.