Abstract:Aiming at the low efficiency of using traditional image recognition methods to diagnose self-explosion faults in aerial images of insulators, and the current image detection method based on deep learning is mostly carried out through the cascade network, which is difficult to directly locate the Self-explosive defect block, a method for directly detecting the Self-explosive defect block is proposed. The number of defective samples is expanded through Generative Adversarial Networks to solve the problem of insufficient defective samples; in addition, the Faster R-CNN detector is improved, and the feature maps obtained through multiple convolutional layers are merged to make shallow features can also be fully utilized, enhancing the ability to locate small-scale targets. Experiments show that when using this strategy to directly detect insulator Self-explosion defects, the missed detection rate and false alarm rate are significantly reduced and a good recognition effect can be achieved.