中国科学院自动化研究所机构知识库

Neural networks always require extensive training samples. However, in some special applications, i.e., insulators in high power grid, it is very hard and costly to collect variety-rich samples. In this study, a synthetic method is proposed to generate segmentation training samples for the insulators. Instead of relying on full-fledged Computer Graphic, this study focuses on the training features of neural networks. Based on this synthetic approach, many kinds of insulators samples including positive, empty and fake ones can be constructed, and their quantities are particularly balanced by an equalization strategy. In order to validate these produced samples, three end-to-end segmentation networks are employed to adapt to the generators in an adversarial training framework. Meanwhile, an improved training strategy is utilized to speed up the convergence. Finally, extensive experiments are executed to further analyze the proposed synthetic method and demonstrate its effectiveness for insulators segmentation.