Generative-assisted multi-stage integrated network: Tackling extreme noise in image denoising

• Novel multi-stage network tackles extreme noise in image denoising effectively. • Proposes a generative-assisted approach integrating diverse functional net-blocks. • Leverages multi-stage architecture to balance noise extraction and detail recovery. • Achieves state-of-the-art results for image denoising across widely-used datasets. • Excels in restoring textures, fine details, color tones under high noise conditions. Image denoising is a fundamental task in computer vision, particularly critical in extreme noise scenarios where noise severely degrades image quality and fine details. Existing denoising approaches often struggle to effectively suppress noise while preserving critical structural and textural details under such conditions. In this work, we propose a generative-assisted multi-stage integrated network (GainNet) designed to address these challenges. The proposed GainNet integrates three key components: the noise extractor block for iterative noise suppression, the image-to-image translator block leveraging conditional generative adversarial networks for direct noisy-to-clean image translation, and the depth-fusion enhancer block, utilizing a swin-convolution architecture to fuse and refine multi-channel inputs. Extensive experiments on widely used datasets (i.e., CBSD68, Kodak24, Urban100, and Set5) demonstrate that GainNet significantly outperforms benchmark models in both PSNR and SSIM, achieving superior noise reduction and texture preservation, particularly under extreme noise conditions (σ ≥ 50). Additionally, to further validate the effectiveness of GainNet beyond standard datasets, a privately curated dataset (AVIP) was introduced for evaluation. The results highlight the ability of GainNet to recover intricate details while maintaining color tones, setting a new standard for denoising in challenging conditions. These findings underscore the potential of GainNet for real-world applications in aerospace imaging, autonomous navigation, and medical diagnostics. The source code and trained models are made publicly available at https://github.com/AVIP-laboratory/Generative-assisted_multi-stage_integrated_network .