Abstract:
Due to the details of low-light images are hard to distinguish, this sort of images is difficult to make further use. To improve the visibility of low-light images, we designed a model to solve the problem of insufficient image feature extraction by traditional U-net. By combining the advantages of deep residual network with strong feature extraction capabilities, we proposed an enhanced algorithm which based on Retinex theory and residual neural network is proposed in this paper. Firstly, we used an improved U-net network with a series of convolutional and up-sampling layers to decompose the image into a reflection part and a lighting part. Secondly, in order to better retain the detailed features, on the one hand, the reflected part and the lighting part obtained from the decomposition are transmitted through a series of convolution blocks, and then sent to the constructed residual network for reconstruction to obtain the restored image, on the other hand, the lighting part is enhanced by four convolution layers to obtain the adjusted illumination map. Finally, the reflection map and illumination map are merged to obtain the final enhanced image. The results show that the improved algorithm effectively improves the visibility of the dark part of the image, and at the same time enhances the color depth and contrast. Compared with other methods, the proposed approach obtains a better performance both in subjective and objective evaluation.