Research And Implementation Of Mage Rain Removal Theory Based On Deep Learning

In recent years,with the continuous breakthrough of hardware computing ability,the application of computer vision has also been developed rapidly.Target detection,target tracking and other technologies have been widely used in intelligent transportation,automatic driving and other fields.However,in the outdoor scene,some bad weather conditions,such as rain and snow,will have a serious impact on the image quality of the camera,resulting in a variety of visual applications can not play its due effect.Therefore,how to overcome the influence of bad weather and restore the damaged image has become a very urgent problem,which is of great significance to the whole field of computer vision.Although the current single image rain removal algorithm has achieved some results,there are still some problems.Firstly,the existing algorithm can not completely restore the image details,and it is more difficult to deal with the long rain stripes.Secondly,in the heavy rain environment,sufficient water vapor and a large number of rain stripes in the air are superimposed and fused to form fog.The existing algorithms rarely consider the rain fog mixture.In addition,it is difficult to obtain the real rain/no rain image pairs,so the current rain removal networks are trained on synthetic data sets,and the generalization ability is insufficient.Therefore,it is still a very challenging problem to remove rain from images.In view of the above problems,the work of this thesis is as follows:firstly,this thesis analyzes the rain streaks model,and proposes a single image rain removal network based on recurrent convolution,which can make better use of the similar internal information between different rain streaks layers,so as to better complete the rain removal work.At the same time,the non local convolution structure is introduced to make the network have excellent performance in dealing with long rain streaks,and performance of the network has reached the advanced level in the field.Secondly,aiming at the case of rain and fog mixing,this thesis analyzes the heavy fog model,and proposes the image modeling under the condition of rain and fog mixing,and then proposes the defogging sub network based on multi-scale structure,combines it with the cyclic defogging network,also verifies the effect of defogging sub network through experiments.Finally,This thesis studies the application of Gan network in image rain removal.Aiming at the problem of real image rain removal,a new generation network of encoding and decoding structure is proposed,which can better capture the rain stripe.Then the loss function is further improved to solve the problem of artifacts in Gan network generated images.Finally,the test is carried out on the real rain images data set collected by this thesis,and the results are very eye-catching.