Interference Line Removal Using A Progressive Model
Kangquan Mo, Pingjian Zhang · 2022
Interference lines in the text image greatly affects the recognition of the text, thus, removing interference lines can improve the text recognition accuracy of the text image and the robustness of the text recognition model. This paper proposes a progressive model for interference line removal based on a recurrent neural network. The model utilizes the cyclic structure of RNN to perform multiple stages of interference line removal, which are responsible for removing part of the interference lines. In each stage, dilated feature extraction module is adopted to gradually extract richer image features. The feature extraction module aggregates dilated convolutions with different dilation rates, which ensures the extraction of receptive field of various scales. To make full use of shallow feature and speed up network training, dense connectivity is added between feature extraction blocks. The experimental results show that the proposed model can effectively remove interference lines in text images, and achieves better performance compared to other models.