Text removal network based on comprehensive loss evaluation

This paper proposes a text removal model, Text Remove Network (TRNet), which achieves an unprecedented clearing effect of picture text. The network uses a jump-connected U-net structure to encode and decode the generator, so as to obtain clearer and sharper texture details of the original image. To solve the problem of color distortion, the generator removes the batch normalization layer and uses ELUs as the activation layer of all convolutional layers. Through the comprehensive loss, which include reconstruction, content, style, total variation, and structural similarity (SSIM) loss, we can clear the image text and preserve the image information of the background, which solves the problem of incomplete text removal and loss of background texture. The local discriminator is used to evaluate the local consistency of the text erasure area. In a text elimination experiment on a synthetic dataset and the ICDAR 2013 dataset, this method had a good effect on foreground text erasure and background authenticity restoration. Experiments on a comprehensive dataset of real documents also showed good results. To achieve targeted removal of sensitive text information on pictures, we collected datasets based on real and synthetic documents, and experimental results were satisfactory. Compared to current and classic algorithms, our text removal algorithm performs best in Image Quality Assessment (IQA).