Low-Light Image Enhancement Based On Modified U-Net

Recent years, researches in low-light image enhancement has done quite a lot and shown great success in real life application. In this paper, a modified U-Net-based method is proposed by combining with Recurrent Residual Convolutional Units (RRCU) and Dilated Convolution. In our method, we achieve higher accuracy by three enhancements. Firstly, replace the basic 3x3 convolution blocks with RRCU. Secondly, replace the 3x3 convolution bottle neck block with multi-ways concatenation. Lastly, replace the max pooling operation with dilated convolution between upper and lower levels. In experiment, the performance of the proposed modified U-Net network is proved to obtain obviously better accuracy than other existing methods in low-light image enhancement.