Real-world Image Noise Reduction through Attention Mechanisms in Generative Adversarial Networks

Abstract

Real-world Images often suffer from various types of noise, decreasing their clarity and usefulness. In order to tackle this problem and obtain the original scene from the image the noises that occurred on the image has to be removed or the image has to be denoising.Recent years have experienced progress in the field of image denoising by using the strengthof Generative Adversarial Networks (GANs). However, the challenge of retaining original image details while reducing noises from real world images is still ongoing. This research proposed an approach to reduce real-world image noises using GANs enhanced with attention mechanisms, specifically channel attention and self-attention mechanisms. The proposed model, ANR-GAN, is specifically designed to concentrate on relevant features of noisy images, resulting in enhanced noise reduction and better preservation of details. The results we obtained from the experiment demonstrate that our model performs better thanthe baseline D-GAN model, with a PSNR of 37.42 and SSIM of 0.9582. These values from the metrics indicate that the combined use of CA and SA mechanisms within the GAN framework significantly enhances denoised image?��?s quality. ANR-GAN model not only reduce noise effectively but also maintain the contents of the original images; this makesthem highly usable for real-world applications. This research contributes in image denoisingfield by providing a method that integrates attention mechanisms into GANs, offering a more robust solution for reduction of noises that occur in real-world images. The results suggest that attention mechanisms are crucial in the enhancement of GAN-based denoising modelperformances.