Detection of Key Structure of Auroral Images Based on Weakly Supervised Learning

Abstract

Weakly supervised learning is of interest and research by many people due to the large savings in labeling costs. To solve the high cost of manual labeling in the research of aurora image detection, an Aurora multi-scale network for aurora image dataset is proposed based on weakly-supervised learning. Firstly, the feature learning mechanism of dynamic hierarchical mimicking is adopted to improve the classification performance of the convolutional neural network based on the aurora image. Then, the multi-scale constraint is imposed on the network through the multi-branch input and output of different sizes. The final output of the auroral image class activation maps with more ideal results, the critical structure detection of auroral images based on imagelevel annotation is realized. Experiments show that the algorithm in this paper can effectively improve the class activation maps results of the auroral image, and has an ideal detection effect on the vital structure of the auroral image.