Image to Image Translation Networks using Perceptual Adversarial Loss Function

Image to image translation based on deep learning models is a subject of immense importance in the disciplines of Artificial Intelligence (AI) and Computer Vision (CV). A variety of traditional tasks such as image colorization, image denoising and image inpainting, are categorized as typical paired image translation tasks. In computer vision, super-resolution regeneration is particularly important field. We proposed an improved algorithm to mitigate the issues that arises during the reconstruction using super resolution based on generative adversarial network. It is difficult to train in reconstruction of results. The generated images and the corresponding ground-truth images should share the same fundamental structure in order to output the required resultant images. The shared basic structure between the input and the corresponding output image is not as optimal as assumed for paired image translation tasks, which can greatly impact the generating model performance. The traditional GAN based model used in image-to-image translation tasks used a pre-trained classification network. The pre-trained networks perform well on the classification tasks compared to image translation tasks because they were trained on features that contribute to better classification. We proposed the perceptual loss based efficient net Generative Adversarial Network (PL-E-GAN) for super resolution tasks. Unlike other state of the art image translation models, the PL-E-GAN offers a generic architecture for image super-resolution tasks. PL-E-GAN is constituted of two convolutional neural networks (CNNs) that are the Generative network and Discriminator network Gn and Dn, respectively. PL-E-GAN employed both the generative adversarial loss and perceptual adversarial loss as objective function to the network. The integration of these loss function undergoes an adversarial training and both the networks Gn and Dn trains alternatively. The feasibility and benefits of the PL-E-GAN over several image translation models are shown in studies and tested on many image-to-image translation tasks