Multi-scale Fusion Low Illumination Image Enhancement Algorithm

Existing image enhancement algorithms operate at the same scale, do not consider the effect of sensory field size on image enhancement performance, the network deep downsampling or upsampling is prone to lose the object contour texture information, and most of them are difficult to balance multiple enhancement goals, such as brightness enhancement, detail restoration and noise suppression. Aiming at these problems, this paper proposes a multiscale multi-stage image enhancement network that incorporates the attention mechanism. The network uses the hybrid attention mechanism in each stage to provide illumination a priori information, mines multi-level features through different scale network branches, and uses denoising network to suppress the noise and color difference of the enhanced image.