Multi-Head Attention of Optical Character Recognition on Relay Protection Drawings
Liangliang Song, Shuyi Zhuang, Chaoyu Gao, Yi Yang · 2023
Recognizing engineering drawings becomes an urgent problem to be solved for digital application in power system relay protection field. The optical character recognition is an automatic technology to recognize characters in images, which is vital for drawing digitalization. However, existing studies on the optical character recognition can neither recognize the irregular shape of texts on the power grid engineering drawings nor accurately identify characters with concentrated distribution. In this paper, we present a new multi-head attention based method leveraging differentiable binarization, which is called EDOCR. The core is the seq2seq model with the multi-head attention strategy to capture the most associated information of the input image for text recognition. Experimental results on the real-world datasets demonstrate the advantages and competitiveness of our method.