Text Recognition in Low Resolution Images Using Trainable Regularization

抄録

In text recognition task, a part of a text in an image often suffer from the low-resolution problem. Consequently, the recognizer cannot predict the character correctly. To address this problem, we present end-to-end text recognition for the low-resolution image in two stages. The image resolution enhancement is applied before performing the recognition process. Our focus in this paper is the modified loss function for the image resolution enhancement. Normally, the super-resolution model traps an overfitting problem, namely, some characters are predicted in another one that has a similar shape. To avoid this overfitting, the regularization term is normally added to the loss function with the fixed weight ratio, which is hard to optimize. In this paper, we make this fixed weight ratio into a trainable parameter that can be optimized in the backpropagation process. We test this approach with many recognizers and we get the improved results. It can achieve the text recognition accuracy of 76.5% in test set and the highest IQA scores.