Abstract:In power line contact tree faults, traces left on the power line surface are an important basis for accident prevention and responsibility determination. However, there is extremely limited research on the feature analysis and identification methods of discharge traces both domestically and internationally. To solve this problem, a 10kV tree-line discharge experiment platform is built to collect power line surface trace images after discharge, and systematically analyze the characteristics of line surface traces, which provide a basic reference for manual inspection and intelligent trace recognition. Then, to solve the problem that the initial frame cannot be automatically determined in the Grabcut algorithm, an improved Grabcut foreground extraction method is proposed, which comprehensively uses the automatic segmentation characteristics of U2Net and the high-precision advantages of Grabcut, to achieve automatic and accurate segmentation of line traces area under complex background. Finally, a comprehensive representation of power line surface traces is proposed based on texture, color feature at low level and deep feature at high level. A majority voting rule is adopted to achieve the decision fusion of the recognition results at low level and high level, and the recognition results of line traces are obtained. The average recognition accuracy rate reaches 91.68% in the test experiments, which proves the effectiveness of the proposed method.