YOLO–DTAD: Dynamic Task Alignment Detection Model for Multicategory Power Defects Image

Abstract

Inspection images of transmission lines from autonomous aerial vehicle (AAV) often contain complex backgrounds and multicategory of power defects. Similarities in categories and scale differences in objects make the traditional single-category detection methods of power defects have unacceptable errors. Therefore, in this article, the you only look once (YOLO)–dynamic task alignment detection (DTAD) model for images of multicategory power defects is constructed to ensure real-time detection of AAV. First, the detection head of DTAD embeds a feature extractor built by grouped convolution in a decoupled head structure of classification and localization, which further learns the task interaction features to improve the model performance. Second, based on the idea of exponential moving average (EMA), the EMA SlideLoss (ESLoss) function is proposed to self-study the intersection over union (IoU) threshold of the bounding box to control the balance between positive and negative samples and dynamically regulate the loss weights of the samples. Finally, normalized Wasserstein distance (NWD) is introduced to alleviate the regression bias of the multiscale object bounding box. Compared with other detectors, the proposed model reaches the mAP50 of 48.2% and 150 frames/s (FPS), respectively, in a private power dataset containing six categories of power defects, which achieves the best tradeoff between speed and accuracy. In addition, generalization experiments are also conducted on other four public datasets to prove the versatility and effectiveness of the proposed model, and the precision values are improved by about 3.1% on average.