A novel image denoising algorithm based on least square generative adversarial network
Mohammed, Sharfuddin Waseem,Murugan, Brindha
DOI: https://doi.org/10.1007/s11554-024-01447-3
IF: 2.293
2024-04-25
Journal of Real-Time Image Processing
Abstract:In recent years, computer vision models have shown a significant improvement in performance on various image analysis tasks. However, these models are not robust against noisy images. There are various causes of noise including noise coming from image capturing conditions such as lighting, weather, and the noise induced by physical measurement devices such as the camera or sensor. To address the problem of noise in images, various works have been proposed. Most of these works focus on computer vision models that run on high-compute devices such as mainframes. However, with recent advances in the deployment of AI technology in mobile and edge devices, there is an increased need for models which work equally well on such devices. In this work, a novel image-denoising algorithm based on the Least Square Generative Adversarial Network (LSGAN) is proposed. It also has the ability to work in settings with lower computing access such as edge devices. The generator part of the LSGAN is a UNet-based architecture which is used to capture detailed low-level features while preserving image information. For the discriminator part of LSGAN, a fully connected network is used, which offers faster convergence, and more stability when training using the composite loss function formed with least-squares loss, visual perception loss and the mean square loss. The proposed model has been evaluated on publicly available denoising datasets including PolyU, CBSD68, DIV2K, and SIDD, under low-compute conditions. The obtained results demonstrate a considerable improvement when compared to various widely deployed works on low-computing devices.
computer science, artificial intelligence,engineering, electrical & electronic,imaging science & photographic technology