An Input-Perceptual Reconstruction Adversarial Network for Paired Image-to-Image Conversion

Abstract

Image-to-image conversion based on deep learning techniques is a topic of interest in the fields of robotics and computer vision. A series of typical tasks, such as applying semantic labels to building photos, edges to photos, and raining to de-raining, can be seen as paired image-to-image conversion problems. In such problems, the image generation network learns from the information in the form of input images. The input images and the corresponding targeted images must share the same basic structure to perfectly generate target-oriented output images. However, the shared basic structure between paired images is not as ideal as assumed, which can significantly affect the output of the generating model. Therefore, we propose a novel Input-Perceptual and Reconstruction Adversarial Network (IP-RAN) as an all-purpose framework for imperfect paired image-to-image conversion problems. We demonstrate, through the experimental results, that our IP-RAN method significantly outperforms the current state-of-the-art techniques.

Adversarial Gaussian Denoiser for Multiple-Level Image Denoising

“…The perceptually inspired denoising method [ 46 ] uses skip-connections in the encoder-decoder network for securing larger receptive fields. However, the skip-connections cause unwanted information flow from the encoder layers to the decoder layers, producing unpleasant images [ 32 ].…”

“…The pixel-wise calculation can produce reasonable images. Though, during some instances, these loss functions mostly catch low frequency rather than high frequency elements of images, resulting in certain critical performance drawbacks (e.g., image artifacts and image blurring) [ 31 , 32 ].…”