RGB–NIR IMAGE DENOISING AND FUSION PREDICTION SYSTEM

Abstract

Digital images captured under low-light, noisy or otherwise unfavourable sensing conditions frequently suffer from degraded visual quality, which limits their usefulness for downstream computer-vision tasks. This paper presents an RGB–Near-Infrared (RGB–NIR) image denoising and fusion prediction system that uses deep learning to remove noise from RGB images and to fuse them with complementary structural information carried by a NIR image, producing a visually enhanced output. The proposed system implements two alternative deep learning pipelines: a two-stage FusionNet pipeline, comprising a dedicated Denoising Subnetwork followed by a Fusion Subnetwork, and an Attention-based Multi-modal Fusion Network (AMFNet) that learns a joint RGB–NIR representation end to end through an attentionfusion block. Input images are decoded, resized, and normalised during preprocessing, and a NIR channel is synthesised automatically from the RGB image whenever a real NIR capture is not supplied. The complete system is deployed as an interactive Flask web application that lets a user upload a noisy RGB image together with optional NIR and ground-truth images, select a model, run inference, and view the reconstructed output together with the Peak Signal-to-Noise Ratio (PSNR) whenever a ground-truth reference is available. A working deployment of the system was used to generate the results reported in this paper: the AMFNet model, evaluated on a 96×96 test image, achieved a PSNR of 25.28 dB together with a visibly denoised and detail-preserving fused output. These results demonstrate that the proposed hybrid denoising–fusion pipeline, combined with a lightweight web-based interface, offers a practical and extensible solution for RGB–NIR image enhancement in research and real-world computer-vision applications.