| Due to the limitations of light,in many cases,pictures are captured in low-light conditions.These picture not only reduce the visual beauty of the image but reduce the performance of many computer vision and multimedia algorithms significantly.Besides,most of the current algorithms are mainly designed for high quality input.Therefore,low light enhancement is also a hot topic in the field of image processing.The challenge of low-light image enhancement includes three aspects: firstly,the paired image of low-light and normal light used to train by deep learning is difficult to obtain;Secondly,recent lowlight enhancement techniques ignore the local characteristics,which are often not well preserved.Third,the existing data sets for low-light image enhancement lack real-time data,and their application value is difficult to be verified.Based on the above problems,this paper puts forward the non-paired low-light image enhancement technology based on GAN which emphasis on local feature preserving.The model is mainly enhanced from three aspects: using non-paired low-light enhancement technology,adding local feature retention part and avoiding the block effect by Sliding window,introducing real-time camera data.The main research points and contributions of this paper are as follows:(1)A low-light image enhancement technique with local feature preservation is proposed to solve the problem of local features neglecting in low light image enhancement.To solve the problem that local features of existing low-light enhancement techniques are not well preserved,this paper put forward an idea to add local perception loss to ensure the local feature to be preserved.This method optimized a particular piece of part each time,and put the processed image block and the corresponding original one into the pretrained network to reserved more low-level features such as color content etc...After many iterations not only local image with More details is preserved but also artifacts are eliminated.Besides,the sliding window can avoid the problem of block effect when processing the local image blocks which improved the quality of the image after augmentation.The experimental results show that the enhanced images in this paper are scored is better than other methods according to the evaluation criteria designed in this paper.(2)This paper introduce a cyclic consistency loss to generative adversarial network and use unpaired images as data set to reduce the difficulty of data set acquisition.It is time consuming and laborious to acquire a large number of paired images in low-light image enhancement.Besides,it is difficult to train low-light image enhancement algorithm with unsupervised,which is difficult to reach the ideal state quickly.Therefore,this paper proposed an unpaired low-light image enhancement algorithm which uses the cycle consistency loss as the supervision condition for training,which saves a lot of manpower and material resources.(3)Introduced a large number of real-time camera data as the training and testing datasets of this paper.In view of the problem that existing low light enhancement algorithms are usually trained under specific datasets and can not obtain good enhancement results on real-time monitoring data.then low light data of real-time camera is introduced to verify the model,which improves the practicability of the model. |
