Transformer Component Recognition in Pictures Taken by Submersible Robots

Inspection of transformers is very important to ensure power supply reliability. Compared with manual methods, using submersible robots to automatically take and analyze pictures is much more time-efficient and lower-cost. To solve the problems of varying illumination and view point in transformer component recognition task, a framework called Transformer Network with Image Enhancement (TRIE) is proposed. This method first enhances the picture based on local contrast information, and then recognizes the inside components with the help of Transformer Network which considers extra context information in the unique encoder-decoder structure and attention mechanism. Experiments show that this framework performs much better than other three deep-learning models on field data of transformer inside pictures, largely improving the development of automatic power equipment inspection.