Power line insulator defect detection using CNN with dense connectivity and efficient attention mechanism

Power line insulator defect detection is an extremely important technology to ensure the safety of power lines. In recent years, electric power enterprises often use UAVs to conduct safety inspections of power lines. This is a resource-limited terminal platform that cannot sustain the huge computational burden. In addition, insulator images taken by UAVs usually have complex background interference. All these require that the power line insulator defect detection algorithm must guarantee high detection accuracy while keeping the computational cost low. To this end, we designed a novel single-stage detection model that can be trained end-to-end based on Yolov3. Our improved model replaces the backbone network of Yolov3 with ResNet50 to reduce the number of model parameters. We changed the original connection structure in ResNet50 to a dense connection to improve the feature extraction capability of the backbone network. To overcome the complex background interference, we add an effective attention mechanism at the end of each layer of the backbone network to enable the model to focus effectively on the detected objects. We also use Mosaic and Random Erasing methods to enhance the dataset. Extensive experimental results show that the model achieves better prediction performance compared to other state-of-the-art methods.