Pix2Pix Generative Adversarial Network with ResNet for Document Image Denoising
Pranjal Jadhav, Sawal Mayuree, Zagade Anushka, Prerna Kamble, Prajakta Deshpande · 2022 4th International Conference on Inventive Research in Computing Applications (ICIRCA) · 2022
Noise degrades quality of scanned document images and adversely affects the accuracy of document digitization and text extraction tasks such as optical character recognition (OCR). Denoising and quality enhancement is the preprocessing stage of the processing pipeline in OCR. This research work proposes an effective end-to-end framework that uses the pre-trained pix2pix Generative Adversarial Network (GAN) to denoise degraded electronic document images. To increase the capacity of the generator network, a variation of the baseline model is developed by replacing the U-net architecture with ResNet6. Along with the discriminator patchGAN in pix2pix model, a pipeline has been developed to extract the patches from input images, predict clean patches using the trained model and finally merge the output patches smoothly. For training, a noisy scanned document dataset created by synthetically adding noises has been utilized to clean the images. Finally, the proposed model is tested by performing quantitative analysis based on different metrics - Structural Similarity Index Measure (SSIM) and Peak Signal to Noise Ratio (PSNR) as well as qualitative analysis by using OCR test on test dataset and real-time documents.