COT2DNet: Contextual 2D Attention for Scene Text Recognition

Although scene text recognition (STR) methods have made great progress, reading the text of irregularly shaped scenes remains a challenge. The current conversion of two-dimensional (2D) images to one-dimensional (1D) feature maps for sequence recognition may lose a large amount of spatial information and cannot recognize text with a large degree of curvature. In this paper, we propose an network infrastructure to recognize texts of arbitrary shapes, named COT2D Attention Text Recognition Network (COT2Net). Firstly, deformable convolution is added to the feature extraction module, which can perform feature extraction according to the shape of the character and expand the receptive field of feature extraction; Finally, in the Transformer encoding stage, the multi-head attention is changed to Contextual 2D Multi-head Attention (COT2D), which can make full use of the relationship between the input keys Q, K, and V to guide the learning of dynamic attention, thereby enhancing the visual expression ability of text images. The COT2DNet uses COT2D to learn the relevance of each character in a scene text image and achieves excellent results on public datasets.