Single Image Derain Based On Dense Convolutional Neural Network

Images taken outdoors in rainy weather are randomly distributed with rain lines in many directions.These dense rain lines obscure the target object and cause light reflection on the background,which in turn reduces the contrast of the target object.Eventually,the details of the shooting are lost and the image quality is not satisfactory.However,The outdoor traffic monitoring and military target reconnaissance rely on clear images,which seriously affect system performance.If we can perform a series of processing operations on the rain line to clarify the image,it will play a crucial role in image recognition,tracking and other fields.At present,single-image rain removal methods are mainly Based on image layer modeling or raindrop detection.These methods either decompose the image,reconstruct the rainless image by sparse coding of the rain map dictionary,or detect the raindrop,use the rain feature to detect the rainline to rain,and further rain the rainline to rain,although all mainstream methods can reach a certain degree of rain effectly,However,different algorithms have the problems of unsatisfactory rain and real-time performance,loss of image details and destruction of image structure.To solve the above problems,in order to improve the effectiveness of the single image rain removal method,The main research work and innovations are as follows:(1)Construct a standard data set for training neural networks by collecting in different outdoor scenes rain-free maps and artificially synthesize a large number of rain maps.The rain map of the data set includes rain lines with different directions and densities under different scenes.Use a convolutional neural network to rain the rainy images in the data set,and then use the corresponding rain-free images.Comparing the images allows the qualitative and quantitative assessment of the effect of rain removal.Accelerate the speed of convergence.(2)Research use of filtering algorithm to extract high-frequency part of the data set to enhance image sparseness,By preprocessing the data set,So as to effectively learn the characteristics of the rainline and improve the training efficiency of the neural network.(3)Research uses a dense convolutional neural network to densely connect the network without deepening the number of network layers,so as to achieve the purpose of re-sampling so as to fully learn the characteristics of rain lines,and further uses a single-layer convolutional neural network to repair the results of the initial raining.The non-linear mapping result of the rain image produces a clear image directly.