Pix2Pix Generative Adversarial Network with ResNet for Document Image Denoising

Revisiting crowd counting: State-of-the-art, trends, and future perspectives

“…The performance is evaluated over PSNR and SSIM metrics. A pretrained Pix2Pix GAN is used in [36] to denoise degraded electronic documents. The architecture replaces the U-net network in the generator with ResNet6, and patchGAN in the discriminator network.…”

Novel hybrid integrated Pix2Pix and WGAN model with Gradient Penalty for binary images denoising

“…Pix2Pix, a conditional GAN architecture, is well-suited for image-to-image translation tasks and has shown effectiveness in denoising applications. By incorporating residual blocks in the bottleneck, as demonstrated in [14], we can mitigate issues related to gradient vanishing and explosion. WGAN-GP introduces a Lipschitz continuity constraint that enhances training stability and reduces susceptibility to mode collapse.…”

“…• Unlike [14] and [15], which are susceptible to mode collapse, a hybrid Pix2Pix WGAN-GP framework is proposed. This innovation aims to enhance the training stability of GANs, addressing the notorious challenges faced during standard GAN training.…”

“…3 illustrates the detailed architecture of this network. Following the approach of [14], a pass consisting of a series of six ResNet blocks using 2D-Convolution is adopted, with Batch Normalization, Leaky ReLU activation functions, 2D-Convolution, and a Batch Normalization layer. This has the advantage of reducing the effect of the vanishing or exploding gradients and in our case, it worked better than a U-NET architecture [46].…”