Abstract: Regarding the issue of missed, erroneous, and false detections in the detection of multiple defects in insulators, an IND-YOLO insulator defect-detection algorithm based on improved YOLOv8 algorithm was proposed. By combining the deformable convolution Dcnv2 and the C2f structure in the network, the algorithm parameters could be reduced, and more attention could be paid to the variable characteristics of insulator defects algorithm. The coord attention attention mechanism was used to improve the feature extraction ability of the region of interest. Considering the insufficient ability of the model to detect small target defects, IND-YOLO algorithm added shallow output. The Siou loss function was used to accelerate the convergence speed of the model. The experiment results show that the average detection accuracy Map@0.5 of the proposed model reaches 0.943, which is 5.6% higher than that of benchmark model. The F1score reaches 0.92, which is 4.6% higher than that of YOLOv8, and the detection speed can reach 62 frames per second. The proposed model has a wide application prospect in insulator multi-defect detection.

Key words: traffic and transportaion engineering; electrified railroad; insulator; YOLOv8 algorithm; deformable convolution; small target

摘要： 针对绝缘子多缺陷检测易发生漏检、错检以及误检等问题。提出一种基于改进YOLOv8算法的IND-YOLO绝缘子缺陷检测算法。通过结合可变形卷积Dcnv2和网络中的C2f结构，可降低算法参数并更注重绝缘子缺陷多变的特征。采用CA(coord attention)注意力机制提升对感兴趣区域的特征提取能力。考虑到模型检测小目标缺陷能力不足的情况，IND-YOLO算法增加了浅层输出。利用Siou损失函数加快模型收敛速度。实验结果表明:该模型检测平均精度Map@50达到0.943，较基准模型提升了5.6%，F1分数（F1score）达到0.92，较YOLOv8算法提升了4.6%，且检测速度FPS可以达到62帧/s。该模型的提出在绝缘子多缺陷检测中具有较广泛的应用前景。

关键词: 交通运输工程；电气化铁路；绝缘子；YOLOv8算法；可变形卷积；小目标