To improve the stability of the bridge structure, we detect bolts in the bridge which cause the symmetry failure of the bridge center. For data acquisition, bolts are small-scale objects under complex background in images, and their feature expression ability is limited. Due to those questions, we propose a new bolt positioning detection based on improved YOLOv5 for bridge structural health monitoring. This paper makes three major contributions. Firstly, according to the calibration anchor boxes of bolts, the size and proportion parameters of the initial anchor boxes are optimized by K-means++ clustering algorithm to solve the initial clustering problem of anchor boxes in object detection. Second, the hypercolumn (HC) technique fuses the low-level global features of the trunk and the high-level local features of three different scales to solve the problem of the inefficient distribution of anchors and insufficient extraction of classification features. In this way, we improve the detection accuracy and speed of bolt detection. Finally, we establish a dataset of bridge bolts through network collection and public datasets, including 1494 images. We compare and verify the new method in the collected bolt dataset. The experimental results show that the precision (P) of the improved YOLOv5x is up to 87.3%, and the average precision (AP) is up to 86.3%, which are 6.5% and 5.9% higher than the original YOLOv5x, respectively.

Synthetic aperture radar (SAR), which can work normally under various meteorological conditions, has been widely researched and applied in marine vessel target monitoring and identification. Among the many research topics, due to the inconsistency of ship scale in SAR images, susceptibility to sea and land noise and clutter interference, resulting in a low detection rate of near-shore ship targets and inaccurate edge delineation of densely lined ships, a target detection algorithm based on deep convolutional neural network is proposed. The algorithm employs the channel-space grouping attention mechanism during feature extraction to enhance features by utilizing global positional and edge information associated with instances. The feature mobility fusion module is employed to merge features of various scales, bolster the interconnection among these features, and enhance multiscale ship target detection capabilities. The decoupled head is employed for ship target localization, while the angle-weighted intersection over union is used to mitigate regression errors. The experimental results show that the precision (P) achieved on HRSID and SSDD datasets reaches 94.81% and 99.01%, respectively, exceeding the control algorithm by more than 1.35% and 0.94%; the average precision (mAP) reaches 92.06% and 99.50%, respectively, exceeding the control algorithm by more than 2.32% and 2.51%; this indicates that the proposed algorithm has a good performance on SAR image ship detection and a strong generalization ability.

RGB–IR object detection exploits complementary texture and thermal cues, but its performance often degrades when modality reliability varies spatially because of illumination changes, occlusion, sensor noise, and ther-mal clutter. Existing fusion methods typically learn deterministic map-pings and may compromise unimodal representations during joint train-ing, while providing limited detection-oriented semantic guidance.We pro-pose MoDiDet, adiffusion-guided framework that jointly enhances unimodal representations, reffnes local modality reliability, and aligns fused features with detection semantics. First, the Modality-Speciffc Distillation Feature Enhancement (MDFE) module transfers knowledge from frozen RGB and IR teachers to preserve the discriminative capability of each branch.Second, the Diffusion-Guided Residual Illumination-Aware Fusion (DRIF) module treats an illumination-derived weight as a coarse reliability prior and pre-dicts a bounded residual correction through conditional denoising, enabling ffne-grained and spatially adaptive fusion. Third, the Cross-Modal Detection-Map Joint Denoising (CMDM) module encodes detection labels, teacher re-sponses, and illumination cues into a uniffed latent map to provide detection-oriented semantic supervision.On DroneVehicle, MoDiDet achieves 83.3% mAP50 and 63.9% mAP50-95, improving the baseline by 4.6 and 5.9 percentage points, respectively.On LLVIP, it obtains 95.8% mAP50 and 69.0% mAP50-95, including a 3.2 percentage point gain in mAP50-95.These results demonstrate improved localization quality and robustness in low-light, occluded, vague and densely populated scenes.