Scene Text Recognition with Linear Constrained Rectification

Scene Text Recognition remains a challenging problem because of various styles and image distortions. This paper proposed an end-to-end trainable model with a rectification module network.The rectification module adopts a polynomial based spatial transform network to rectify the distorted input image, the feature representation between the rectification and encoding step is shared. The model can be trained with the scene images and the corresponding word labels. With the flexible rectifying and feature sharing, this model outperforms previous works through the extensive evaluation results on the standard benchmarks, especially on irregular datasets, 80.2% on IC15 and 85.4% on CUTE, more specifically.