Research On Rain Removal Method Of Single Image Based On Deep Neural Network

The single image rain removal task is an important part in the field of computer vision tasks,and plays an important role in vision tasks such as target tracking and unmanned driving.Therefore,in a variety of computer vision tasks,how to efficiently obtain a better rain removal effect in a single image has become a key problem in the single image rain removal task,and it is also an urgent problem in the field of computer vision.Through in-depth analysis of the research status of single image rain removal methods at home and abroad,combined with deep learning ideas and computer vision related knowledge,we have conducted in-depth research on how to effectively obtain better rain removal effects in a single image.Firstly,the three major characteristics of rain are introduced,and the convolutional neural network involved in this subject and its basic knowledge and commonly used recurrent neural network are introduced accordingly.Secondly,a single image de-raining method based on guided filtering and image enhancement and using deep learning ideas is proposed.This method firstly decomposes the rainy image into a smooth base layer and a high-frequency detail layer based on guided filtering;secondly,it proposes an adaptive Gamma correction algorithm to enhance the smooth base layer to improve the contrast;again,constructing a fusion residual block and channel attention mechanism The deep neural network realizes the high-frequency detail layer to remove rain;finally,the enhanced smooth basic layer is merged with the high-frequency detail layer after rain to achieve single image rain removal.Thirdly,a single image rain removal method based on convolutional neural network and recurrent neural network combined with recursive structure is proposed.This method first designs the separable calibration module and separable residual block based on the dual attention mechanism to obtain the calibrated deep feature map;secondly,the deep feature map is input into the GRU cycle unit that can make full use of the recursive structure to obtain the stage Feature map;again,the stage feature map is used to obtain the rain image of the current stage through dimensionality reduction;finally,the rain image of the current stage is re-entered into the network as the rain image of the next stage,and the final result can be obtained after recursive multiple times De-rain image realizes single image de-rain.Finally,the corresponding experimental verification and analysis are carried out for the two mentioned methods of rain removal.