A Flow-Appearance Fusion Defect Detection Method for Substation Equipment in Multi-Scene

In the daily maintenance process of substation, the defect detection of substation equipment is a critical problem. This task is challenging considering of the variety of defects, the rarely sample of various defects and the complex environment. This paper presents a flow-appearance fusion defect detection method for substation equipment in multiple scenes, which contains a Siamese optical flow module and a defect location module. The Siamese optical flow module is a pretrained optical flow network, which is utilized to obtain forward and backward optical flow between the normal image and the detect image. The optical flow represents the common changing features of different types of defects in multiple scenes. The defect location module utilizes MI-darknet53 to realize the extraction and fusion of appearance and flow features. Then the extracted multiscale features are used to build the feature pyramid network (FPN), which combines semantic defect information in high level and geometric defect information in low level. Finally, FPN is utilized to predict object-level defect areas in three different scales. Experiments show that our proposed method has good reliability and superior performance in the defect detection for the dataset of multi-scene substation equipment.