Text Detection of Transformer Based on Deep Learning Algorithm

Transformers are important equipment in the power system. At present, the text information collection of transformer nameplates is through manual, which is inefficient. Therefore, it is necessary to find a high-precision automatic detection method of transformer text information. However, the current text detection algorithms have limited ability to detect special characters on the transformer. And they will also have the problem of incomplete detection in detecting the dense text and long text on the transformer nameplate. We propose a text detection network based on segmentation to automatically calibrate the text box of transformer nameplates. Our network is based on DB (differential binarization) network. It has a new feature fusion structure, which refers to the feature fusion structure of the u-net network. The proposed network has achieved better performance than the advanced scene text detection algorithms (DB, East) on the English scene text dataset icdar2015 and the Chinese-English mixed scene text dataset icdar2017. And it also has good performance in GPU occupancy, reasoning speed, and other indicators. The text detection results of actual transformer pictures show that the proposed algorithm solves the problem of poor detection performance of existing deep learning networks in dense text and long text of transformer pictures.