Zero-Shot Underexposed Surveillance Image Enhancement Based on Convolutional Neural Network

This paper proposes a convolutional neural network-based zero-shot underexposed monitoring image enhancement method aimed at addressing the need for underexposed monitoring image correction. The method designs a zero-shot convolutional Retinex decomposition network (ZSCRDNet) and constructs an underexposed monitoring image multiple exposure frame (UMIMEF) based on it. ZSCRDNet is based on the retinal cortical theory and uses a convolutional neural network as an image processing tool to decompose underexposed monitoring images into three components: illumination, reflectance, and noise, through zero-shot learning. Finally, the image is enhanced by subtracting the noise and enhancing the illumination component. UMIMEF achieves multiple enhancements of underexposed monitoring images through residual structures and inhibition of single-time enhancement, enhancing monitoring images of varying degrees of underexposure to sufficient intensity. To evaluate the performance of this method, we collected a dataset of unpaired underexposed monitoring images (UUMImages) in a simulated underexposed monitoring image environment, totaling 127 images. Extensive experimental results show that the enhancement results of the method outperform the existing image enhancement methods in both PIQE and BRISQUE evaluation metrics, and the visual effects are generally better. This indicates that the method can effectively improve the quality and visual effect of underexposed surveillance images.