A real-time text recognition based on involution operators and graph convolutional networks
摘要
The ambiguity and stickiness of text in natural scenarios make text recognition models incorrectly recognize certain characters as others. To address the above problems and the computational scale of the model, we propose a real-time effective text recognition method (IGTR). The method introduces the involution operator in extracting feature sequences within a larger sensory field using a smaller number of parameters. Due to the dependency of the characters before and after the text line, after obtaining the text feature sequence, IGTR uses the gate recurrent unit (GRU) algorithm to obtain richer contextual semantic information and pairs it with the graph convolutional neural network (GCN) to further enhance the semantics of the text features on the basis of the structure of the graph. Finally, the CTC algorithm is used to optimize the computation process of the loss function, and the dynamic planning of the search path is adopted to further reduce the computation amount of the text recognition model. Comparative experiments on ICDAR 2013, ICDAR 2015 and SVT datasets show that the method can effectively recognize multi-scale fuzzy text and sticky text within 75ms. With no dictionary supervision, the