Low-Light Image Enhancement Network Based on Multiscale Interlayer Guidance and Reflection Component Fusion

Images captured under the influence of external factors (such as low light, nighttime, complex weather conditions, etc.) often exhibit unpleasant visual effects. Previous image enhancement methods have overly focused on improving brightness, neglecting the preservation and enhancement of image detail and color features. Therefore, this paper proposes a network with multi-scale interlayer guidance and reflection component fusion (defined as MGRF-Net) is proposed for low-light image enhancement. Among them, the reflection component is obtained from the decomposition sub-network by Retinex decomposition, and is simultaneously enhanced with the low-light image through the multiscale interlayer guidance sub-network, so as to obtain the clear and convergent illuminance estimation and the low-noise reflection component, and finally the two are fused to obtain the final enhanced image. Specifically, the multi-scale inter-layer guidance sub-network introduces three efficient fusion feature modules: the feature guided enhancement module, the feature learning module, and the feature cross-learning module. These modules are respectively used to extract the underlying feature information to guide the upper layer of features for detail enhancement, enhance and converge the guided features of each layer, and preserve the skip connection and up-sampling features in the U-Net structure. Additionally, three feature extraction modules are designed: spatial-channel attention, global feature-extraction block, and multi-scale extraction block to extract local and global features. Experimental results show that the proposed method outperforms other advanced methods in both visual effects and quantitative aspects.