Defective Insulator Detection Algorithm Based on Improved YOLO V7 Lightweight Model

In order to address the storage limitations of the UAV host and to overcome the conflict between detection speed and accuracy of previous models, we designed a lightweight network model inspired by YOLO (You Only Look Once) v7. However, this task was initially hampered by the paucity of images describing the defective components. To address this problem, we introduced S-DDIM, a modification of DDIM that generates optimally sized datasets for training and evaluation purposes. Moreover, we have engineered an innovative convolutional kernel named BODconv. This kernel is designed to supplant the conventional convolutional module. It is integrated with a cutting-edge attention-augmented detection head, culminating in the BODconv-D detection framework. This integration markedly amplifies the network's prowess in feature extraction and enhances its learning efficacy. Furthermore, by utilising a parameter-dependent <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\gamma$</tex>-factor channel pruning technique, we significantly reduce the model size and inference time. We perform a comprehensive comparative analysis using an open source dataset and an extended insulator failure dataset. The results show that the improved algorithm proposed in this study significantly improves the average accuracy by 4.4 %, dramatically accelerates the detection speed by 28.9 %, and dramatically reduces the model size by 57.7 % when compared to the original YOLO v7 method.