Multi-Scale Feature Fusion for Low-light Image Enhancement in the RAW Domain

With advancements in technology, low-light image enhancement has become a popular research topic. Traditional methods are constrained by fixed processing windows or specific algorithm frameworks, making it challenging to fully capture and enhance global details in low-light images. Deep learning models based on self-attention mechanisms excel at capturing long-range dependencies but come with high computational costs and limited ability to enhance local details. To address these issues, we propose a Multi-scale Feature Fusion Network(MFFNet) that integrates self-attention mechanisms with convolutional neural networks(CNN). By employing local and global feature extraction modules, MFFNet effectively captures features at different semantic scales, enhancing global luminance distribution while preserving rich local details for more refined and comprehensive enhancement. To further enhance the quality of image recovery, we design the discriminator module to distinguish the generated image from the real image, thus providing additional supervisory signals. The discriminator is able to learn the subtle differences between the real image and the recovered image and guide the generative model to produce more natural and realistic enhanced images.