Abstract: Aiming at the problems of large workload and low intelligence of the current infrared image-based overheating defect detection techniques for composite insulators, and the poor accuracy and poor generalization performance of the traditional image segmentation methods in complex backgrounds, an overheating defect detection method is proposed for composite insulators based on instance segmentation network Mask R-CNN. Firstly, in order to improve the accuracy of segmentation, the Mask R-CNN network is improved according to the idea of Cascade R-CNN, and the data augmentation and transfer learning methods are used for model training to improve the network performance. Secondly, the result obtained by deep segmentation network is further optimized by using traditional image processing methods such as skeletonization, so that the final segmentation result only covers the core rod of the composite insulators. Finally, the temperature data in the infrared image is directly read and converted into the actual temperature value, and the grade of overheating defects is judged according to the relevant methods and criteria provided in DL / T664—2016 Infrared Diagnostic Application Specification for Live Equipment. The results show that the algorithm proposed in this paper has a high detection accuracy of 100% for the infrared images of composite insulators with serious and urgent defects, but has false detection occurrence for the infrared images without overheating defects or with general defects. On the whole, the accuracy rate of 93% is achieved in defect detection of test sets.