Abstract: Aiming at the problems of the large number of aerial image samples acquired by drones, the large background noise of aerial images, and the small number of failures of the grading ring skew, an improved lightweight Mask RCNN feature extraction network is proposed to obtain the contours of the grading ring and the insulator; when judging the skew of the grading ring, the Hough transform is first used for the horizontal transformation of the segmented insulator string, and then the slope of the two points of the grading ring and the edge of the insulator is calculated to judge whether the grading ring is skewed. Experimental results show that under the influence of the basically unchanged accuracy of the contour extraction of the average pressure ring feature, the detection speed of a single picture is reduced by 80 ms compared to the original residual network.