Improved Image Denoising: A Combination Method Using Multiscale Contextual Fusion and Recursive Learning

Abstract

The exponential growth of imaging technology has led to a surge in visual content creation, necessitating advanced image denoising algorithms. Conventional methods, which frequently rely on predefined rules and filters, are inadequate for managing intricate noise patterns while maintaining image features. In order to tackle the issue of real-world image denoising, we investigate and integrate a new novel technique named recursive context fusion network (RCFNet) employing a deep convolutional neural network, demonstrating superior performance compared to current state-of-the-art approaches. RCFNet consists of a coarse feature extraction module and a reconstruction unit, where the former provides a broad contextual understanding and the latter refines the denoising output by preserving spatial and contextual details. Deep CNN learns features instead of using conventional methods, allowing us to improve and refine images. Dual attention units (DUs), in conjunction with the multi-scale resizing Block (MSRB) and selective kernel feature fusion (SKFF), are incorporated into the network to ensure efficient and reliable feature extraction. To demonstrate the advantages and challenges of combining many configurations into a single pipeline, we take a more detailed look at the results. By leveraging the complementary properties of these networks and computational models, we prefer to contribute to the creation of techniques that enhance image restoration while preserving crucial information, therefore encouraging further research and applications in image processing and artificial intelligence. The RCFNet achieves a high structural similarity index (SSIM) of 0.98 and a peak signal-to-noise ratio (PSNR) of 43.4 dB, outperforming many state-of-the-art methods on two benchmark datasets (DND and SIDD) and demonstrating its superior real-world image denoising ability.