Image Quality Restoration Technology Based On Deep Learning

Image quality restoration is one of the basic tasks of computer vision.Images and its accompanying visual algorithms are widely used in classification,detection,monitoring,automatic driving and so on.For example,the image signal received by the camera is often accompanied with noise or bad weather and other adverse effects,which often cause classification errors and inaccurate recognition for subsequent applications.The semanstic information of background images can not be fully mined with the pollution.Therefore,in order to improve the performance of computer vision tasks,image quality recovery has become an essential pre step of image analysis.In recent years,many models based on deep learning have been proposed,which have achieved remarkable results in image quality restoration.These models are built based on the prior of pollution information.Because different pollution has different prior information,most of the models often focus on one of the pollution,but this greatly reduces the generalization ability of the model.For example,the rainy weather is often accompanied by haze,and the noise is widespread in the collection of pictures.At the same time,because it is difficult to collect the paired data sets of bad weather and clean background pictures,the existing picture quality recovery training data sets are designed manually,so the distribution of the pollution is single.The training dataset determines the upper bound of the model's ability,which greatly limits the model performance.Finally,because the loss function often uses the mean square error,it will lead to over smooth restoration of the picture and lose the details in pictures.In order to address these problems,improve the generalization ability of the model,and make up for the details of recovery,in the aspect of dataset construction,this paper designs a training dataset with extensive pollution distribution,including different intensity of rain and fog to cover different distribution of pollution.In the aspect of model construction,we design the pollution intensity perception module and image quality recovery network.The pollution intensity perception module aims to extract the intensity characteristics of pollution,so as to assist the distribution of background pictures.The image quality recovery network is divided into three parts,namely,defogging,rain / noise removal and detail recovery modules.The detail recovery module is based on the Generative Adversarial Network,which can capture the data distribution of the real background image,so as to supplement the details of the generated image.We experimental show that the enhancement effect of intensity aware information on image quality restoration task,and demonstrate the generalization ability of the new model to deal with a large number of different pollution.Compared with many state-of-the-art methods,the new model not only eliminates the pollution,but also takes into account the details of the recovery images.