Research on few-shot power detection of siamese network based on improved RPN

Abstract

In order to solve the problems of difficulty, low efficiency, and insufficient data to support large-scale training in existing power system detection methods, a few-shot detection method based on siamese network was proposed. Firstly, under the framework of Faster RCNN(region convolutional neural network) object detection algorithm, a siamese network model supporting imageand querying image sharing was built. Then, the improved RPN (region proposal network) module was used to generate proposals of higher quality. Finallly, the RoI(region of interest) supporting and querying images was correlated and matched on the detection head. The results show that the proposed algorithm, applied to the self-constructed EPD(electric power detection) dataset, can detect foreign matters in bird nest and insulator in the power background, and the detection index reaches 1892% mAP, in the case of only 10 supporting images. Compared with other algorithms, the siamese network model with small sample size has better performance under extremely few shot situations, and has the advantage of being more lightweight, which provides some reference for the new research direction of electric power detection.

Keywords

• computer awareness; siamese network; electric power; few shot; object detection