Research And Implementation Of Low-light Image Enhancement Algorithm Based On U-Net Network

Low-light image refers to the image collected in low-light environment.For example,the images captured by the monitoring system on the street at night,and the images obtained by the workers in order to survey during the underground coal mining.Low-light image with low SNR and lose texture details,so the processing and analysis of low-light image have always been a challenging task.To enhance the low light image and convert it to the normal brightness image is beneficial to the downstream image processing task,such as target detection,object tracking,image classification,etc.The existing methods of using deep learning to enhance the low-light image are mainly for the images collected in the ordinary low-light scene.For the images with serious underexposure,these methods often do not get the ideal enhancement effect.The content of this research is image enhancement in an extreme low-light environment.In order to effectively suppress image noise,improve the quality of image enhancement,and solve the problem of image enhancement in an extremely low-light environment,this paper introduces wavelet transform and content perceptual loss based on U-Net.The innovation of this paper is as follows:(1)U-net based on wavelet transform.Wavelet transform can highlight the low-frequency information and high-frequency information of the image,which has certain advantages in image enhancement and can use a small amount of information to represent the main features of the image.In addition,considering that both the down-sampling and the wavelet decomposition can make the image size smaller,and both the up-sampling and the wavelet reconstruction can make the image size larger,so in the U-Net architecture,the down-sampling and the up-sampling operations of the corresponding region are replaced by the wavelet decomposition and the wavelet reconstruction respectively.(2)Image enhancement based on content perceptual loss.At present,most of the deep learning methods use the error between image pixels,and lack the difference in perception of image content.Therefore,in the process of network model training,this paper selects the perceptual loss function based on VGG16,selects the feature map of different Relu layers in the network to calculate the loss and explores its effect,so as to further improve the image enhancement quality.In this paper,we use the open-source dataset SID,which contains short exposure and the corresponding long exposure image.For the enhanced image,two indexes,peak signal-to-noise ratio and structural similarity are used to evaluate and analyze from the perspective of visual perception.The experimental results show that compared with the traditional thread processing method and the original U-Net algorithm,the method proposed in this paper can better restore the image texture details,ensure the normal color conversion of the image,and effectively suppress the noise.And through experiments,we can know that the results obtained by using content perceptual loss are more consistent with human visual perception.