Detection of Bolt Defects on Transmission Lines Based on Multi-Scale YOLOv7

Bolts are indispensable in power systems, serving a critical function in securely fastening various components and ensuring stable operation of transmission lines. To address the challenges associated with detecting small bolt objects in images captured by unmanned aerial vehicles (UAVs), we propose an enhanced multi-scale YOLOv7 model called multi-scale YOLOv7 (MS-YOLOv7). First, a multi-scale path aggregation network (M-PANet) is developed, which seamlessly integrates multi-scale information and features four detection heads, thereby reducing the miss rate of small objects. Second, a channel-coordinate attention (CCA) mechanism is introduced to enhance the network's focus on spatial and channel information, thereby augmenting object localization accuracy. Furthermore, in the feature fusion process, the selected convolutions are replaced with a fully self-calibrated convolutional network (FSCNet) to expand the receptive field and enable richer feature extraction. Finally, standard max pooling is replaced with space-to-depth convolution (SPD-Conv) to preserve object detail information during the downsampling phase. Experimental results on a proprietary bolt defect dataset demonstrated significant improvements over the original YOLOv7 model, with a 2.52% increase in mean Average Precision (mAP), a 3.07% improvement in recall, a 1.54% improvement in accuracy, and a 2.41% increase in F1 score. The algorithm also exhibited robust performance on the CSUST Chinese traffic sign detection benchmark (CCTSDB) dataset, demonstrating its strong generalization ability and robustness.