Research On Image Enhancement Algorithm Of Mobile Phone Image Based On GAN

Smartphone's camera is very convenient,is closely connected with People's Daily life,with the development of science and technology and the continuous improvement of people's living standard,people has higher requirement for the quality of mobile phones,but the smaller size of smart phones,bound to lead to the physical limits of built-in cameras,lens optical parameters,the size of the sensor and lack of specific hardware support,etc.,greatly affect the quality of the pictures,can't achieve the SLR camera photo quality.Physical limitations cannot be completely overcome,but image enhancement can be achieved through post processing to improve photo quality.At present,the method of post processing is usually manually adjusted by professional retouching software,which is time-consuming and laborious and requires certain professional skills.Therefore,this thesis proposes a full-automatic end-to-end deep learning method,whose goal is to use the neural network to directly learn the mapping of mobile phone photos to SLR photos.The common deep learning method for image to image conversion is to use the mean square error loss function and convolutional neural networks.The main work of the paper includes the following aspects:(1)Optimize the loss function and network architecture,adding the adversarial loss and content loss to the traditional mean square error loss function to form a multiple loss furnction makes the generated image more realistic,replace the standard convolutional neural networks for the generative adversarial networks,and the introduction of residual network jump connection in it make the image detail better retain information.(2)To develop a mobile phone photo enhancement system based on the Web,transplanting the trained network model to the server to provide a service,the user can upload pictures to the server via the Web client,then the server invocation model to enhance processing,finally converts it to the quality of the SLR photography photo shows to the Web client.Training on a data sets contains thousands of mobile phone/SLR photos,the experimental results show that using the method of paper enhanced mobile photos in clarity,brightness,contrast,color rendering and texture details,etc,have been improved significantly,in both the subjective visual effect and objective evaluation indexes are better than using software adjustment and in recent years,some relevant modification based on the deep learning method is better.