Research On Application Of Attentive Generative Adversarial Network In Raindrop Removal Of Single Image

Rainy day is a common severe weather,the image affected by raindrops will produce serious degradation,resulting the reduction of the recognition,tracking and other performance of the computer vision system.Generative Adversarial Networks has strong ability of feature learning and feature expression.Combining it with attention mechanism,it can generate higher quality images according to the area concerned by attention mechanism and the idea of antagonistic training,which provides a new technology and means for the application of computer vision.It has achieved good results in image raindrop removal.Removing the influence of raindrops and restoring the background scene will facilitate the subsequent image processing.Therefore,the main contents of this thesis are as follows:First,this thesis analyzes the raindrop removal method based on the attention-generating adversarial network.The method combines the attention mechanism with the generating adversarial network.When removing raindrops,the raindrop image is subtracted from the clean image to guide the circulation network to generate raindrop attention maps.Attention is focused on the restoration of the raindrop area.It only focuses on the recovery of raindrop area and ignores the edge structure of background object in raindrop image,which leads to the blurred edge details after rain removal.Aiming at the shortage,this thesis uses the guided image filtering to obtain the raindrop region and edge information,uses the recurrent neural network to guide the raindrop region to generate the raindrop attention map,adds the raindrop attention map to the generating adversarial network model,and introduces the loss of feature space and pixel space in the model.The loss to adjust the details of the raindrop removal effect makes the generated image clearer.In the case of dense and sparse raindrops,the method is compared with other mainstream rain-removal methods,and the experimental results are analyzed and compared.Experiments show that the optimized rain removal method has a good effect on visual effects and objective indicators.Then,the optimized raindrop removal method is applied to the video image to remove raindrops.Removing raindrops from a video image is essentially a single image removing raindrops.Considering the rain removal efficiency of video and the small difference between adjacent frames of video.After video is divided into frames,only odd frames are operated in this thesis.Firstly,the video frame is compressed by bicubic interp,and then the raindrop is removed by the method of attention generating adversarial network after optimization.The experimental results show that the method proposed in this thesis has a better effect on the removal of raindrops when applied to the raindrop video taken in the window.