ETIDCNet:Efficient Text Image Deblurring Convolution Neural Network

In order to restore blurred document images with unknown degradation, many blind deblurring algorithms for single document images have been proposed. However, the evaluation of these algorithms mainly used lower resolution images. Therefore, it is not clear how well these algorithms perform on real-world larger blurred text images and how we can measure progress in the field. In this paper, our goal is to advance the research and application of text image deblurring, we propose a new end-to-end network ETIDCNet (Efficient text image deblurring convolution neural network), which can remove the blur of high-resolution images. Specifically, we designed two modules: EMAM (Efficient mixed attention block) and 3DAM (Simplified Three-dimensional attention block) to enhance ETIDCNet's capabilities. Qualitative and quantitative experiments show that ETIDCNet can solve the text image blur task well. In addition, we construct a TIDD data set containing 1000 high-resolution text images from multiple fields such as information technology, social science, agricultural science, medicine and health, and engineering technology.