A theoretical justification of warping generation for dewarping using CNN

Abstract

Dewarping is a necessary preprocessing step to recognize text from a distorted camera captured document image. According to recent literature, deep learning-based approaches perform with higher accuracy in similar domains. The deep learning-based neural networks are not yet fully explored in the domain of dewarping. To fill this gap, we propose a dewarping approach based on the convolutional neural network. A large number of images are required to train such networks. However, it is a tedious job to capture such a large number of images. Hence, it is required to generate synthetic warped images for the training phase of the deep learning-based neural network. The existing synthetic warped image generation methods are heuristic-based. In this paper, we propose a novel mathematical model for the generation of warped images. The proposed model takes some parameters such as depth of the surface, camera angle, and camera position and generates the corresponding warped image. These parameters are the ground truth for that particular warped image. We use a Convolutional Neural Network (CNN) based model to estimate the warping parameters from a 2D warped image for dewarping. In the training phase of CNN based model, the synthetic images and their corresponding ground truth are used. Next, the trained model is used to dewarp the unknown warped images. The performance of the proposed warping model is analyzed. Finally, the proposed dewarping method is compared with existing approaches. In both cases, the results are encouraging.