Universal Framework of Seals Erasing with Generative Adversarial Network

Abstract

In last decades, methods based on deep learning has dominated many application fields, such as image inpainting, image super high-resolution and image denoising. In the Optical Character Recognition (OCR), we need to erase all seals due to some special reason like improving the performance of invoices and receipts recognition. However, owing to the diversity of the shape, color and position of the seals, it is very difficult to erase all seals without affecting the original information of the invoices and receipts. At present, there are two common methods to erase seals, one is from the perspective of computer graphics, the other is from the perspective of computer software, like PhotoShop. Nevertheless, the common weakness is the poor robustness and inefficient. So, In this paper, We propose an end-to-end network for erasing seals, and we have four contributions. (1) We first propose a universal framework to do the seals erasing by using Generative Adversarial Network (GAN). (2) Training the seals erasing network by the images synthesised by scripts and seals crawled on the website, instead of spending high cost to collect images by ourselves. (3) In terms of speed, due to using end-to-end predict script, using our method to erase all seals on the 512 * 512 RGB image with padding is only 800 ms, much faster than the traditional methods. (4) There is also improvement on the invoices and receipts with the same parameter on the Convolution Recurrent Neural Network (CRNN), the average accuracy of the sequences by using our method is 91.3% higher than the original images is 89.7% on the 800 quota invoice test images.