Research On Low-illumination Image Enhancement Based On Convolution Neural Network

Because of weak lighting conditions or devices with poor filling flash,it is easily to produce low-illumination images when taking pictures.These images are useless because of the large noise and hard recognition.Therefore,the studies of low illumination image enhancement are significant.In recent years,with the spread of deep learning,deep learning based low-illumination image enhancement method,especially the convolution neural network,has gradually replaced the traditional method and become a research hotspot.Therefore,this paper focuses on the low illumination image enhancement method based on the convolution neural network.Due to the lack of an appropriate no-reference image quality assessment in the field of low-illumination image enhancement,this paper proposes a no-reference image quality assessment based on convolution neural network firstly.This method assesses the quality of enhanced images by noise、clarity and visibility.Using the convolution neural network as the feature extractor of the distorted image,and the score will be given by comparing the extracted features.For any input image,the network will output scores of three dimensions.The image quality score of the enhanced image can be obtained by weighted sum of these three scores.Some exiting methods are compared with the method proposed.The experimental results show that the method proposed is more effective.Then,this paper proposes a low-illumination image enhancement method based on convolution neural network.This method is composed of multi-task convolution neural network method and generative adversarial network method.And the multi-task convolution neural network method includes image decomposition using Decompose-Net and image enhancement using Enhance-Net.The Decompose-Net is designed based on Retinex theory,which can decompose the input low-illumination image into illumination and reflection.The Enhance-Net further processes and enhances the decomposed image.At the same time,a Discriminate-Net is proposed to distinguish the enhanced images from the real light images.The Enhance-Net and the Discriminate-Net are composed to the generative adversarial network,and the Discriminate-Net is used to supervise the Enhance-Net for obtaining the better output results.Finally,using LOL dataset 、 Multi-exposure image dataset and real low-illumination images,this method is compared with the existing related methods qualitatively and quantitatively.Analysis through comparison,it can be found that this method has better enhancement effect.The images are less noise and recover more details after enhancement.Also,this method has the better generalization ability and satisfactory processing speed.