Research On Single Image Dehazing Method Based On CycleGAN

In recent years,with the rapid development of big data technology and artificial intelligence technology,new information processing systems,such as vehicle driverless system and traffic intelligent monitoring system,need clearer input images.However,due to the impact of air suspension(such as fog,haze,dust,rain,snow,etc.)in the real environment,the collected images will have serious degradation,such as tone shift,contrast reduction,overexposure,information occlusion and other problems.These problems not only affect people's direct senses,but also bring serious interference to image detection,tracking and recognition by intelligent information processing system.Therefore,it has great application value of image dehazing and visible restoration.Although researchers have made great progress in image dehazing,especially the dehazing methods are based on deep learning have achieved very good results,but these methods are based on the synthetic hazy image as training data,and the effect of dehazing in real environment is not very ideal.The dehaze method based on CycleGAN can solve the problem of data set,but its effect of dehazing is very bad.So far,there are still great challenges in the balance of practicability and accuracy of the existing dehazing methods.(1)In this paper,we propose a dark channel prior cycle dehazing network(DCP cycle dehaze),which is single image dehazing method.The network based on the CycleGAN,added the DCP loss based on the prior knowledge of dark channel and the improved cyclic perceptionloss.DCP-Cycle-Dehaze enhances the ability of the model in the direction of dehazing by increasing the targeted loss function and enhancing the sensitivity of the network to the hazy features in training.It further improves the performance of the CycleGAN framework in the image dehazing task,and make the model achieve the same accuracy as the supervised training in the unsupervised training mode.Through the simulation experiments on four different types of datasets,O-HAZE,I-HAZE,RESIDE and D-Hazy,it is proved that the dehazing network proposed in this paper has achieved good results in the outdoor environment;at the same time,it also has good results in the indoor environment.The experimental results prove the effectiveness of our method from both quantitative and qualitative perspectives.(2)In this paper,we propose a Semi training color stripping DehazeNet(STCSDN).In this paper,a new adaptive dehazing method is proposed,which mainly depends on two important properties of convolution neural network in the process of dehazing.One is that the learning speed of convolution neural network for contour and shadow information is faster than that of color information;the other is that sketch is not sensitive to the haze concentration.Based on property one,STCSDN uses semi-trained generator as sketch module through CycleGAN.The module can extract the haze less gray image from the hazy color image,and the gray-scale image only contains contour and shadow information,discards the original color information of image disturbed by hazy information.This method has strong adaptability,visibility and authenticity,and can be applied to any scene.Further,according to the dehaze task with color demand,another CycleGAN is used to finish the color task of gray image and get the color haze less image.In the process,based on the property two,we can complete the training of painting module only through the sketch extracted from the haze less image without the participation of the hazy image.In addition,due to the natural one-to-one correspondence between sketch and hazy image,the methods that rely on pair data can also be applied to engineering practice,greatly increasing the practicability of these methods.Through the simulation experiments on different types of image dehazing datasets,it is proved that the STCSDN's color stripping idea proposed in this paper can remove the influence of haze,restore image details,and greatly enhance the visual effect.At the same time,the simulation results also prove that STCSDN can combine the existing image conversion model to restore the color and get the real color defog image.