A convolutional neural network based method for low-illumination image enhancement

Abstract

Nowadays, images can be conveniently captured by various image acquisition devices. Weak lighting conditions and devices with poor filling flash will produce low-illumination images. These degraded images are difficult to identify, and must be processed by some methods through the computer. With the inspiring performance of convolutional neural network (CNN) in image classification, object detection and tracking, some studies have been made to enhance low-illumination images by using CNN in recent years. In this paper, based on the existing researches of CNN based low-illumination image enhancement, an improved Unet model is proposed to enhance low-illumination images. At the same time, this paper introduces two new loss functions: Peak signal-to-noise ratio (PSNR) loss and multi-scale Structural similarity (MS-SSIM) loss, and use a mixture of these two loss functions as loss function in our model. Our method can effectively balance the brightness of the processed image, accurately restore the color, so that the enhanced image have a better perception. Results demonstrate that the proposed method outperforms other enhancement methods.