Unconstrained end-to-end text reading with feature rectification

We propose an end-to-end trainable network that can simultaneously localize and recognize irregular text from images. Specifically, we find the feature incompatibility problem, which arises from the contradiction between detection and recognition tasks for feature extraction of the convolutional neural network, and propose to introduce the larger-scale features for the recognition part to improve the accuracy of recognition instead of using the same feature with the detection. To extract effective text features for perspective and curved text recognition, we propose a position-sensitive network to rectify the text proposal features in the recognition branch. The position-sensitive network, which is trained in a weak supervision way, takes the proposal detection feature as input and outputs the feature rectification information. Experiments demonstrate that the proposed method can achieve state-of-the-art or highly competitive performance compared with baselines on a number of benchmarks. (c) 2021 Elsevier B.V. All rights reserved.