Research On Target De-smog Algorithm Based On The Visible Light Polarization Smog Image

In recent years,with the rapid development of the visible light polarization detection technology,the visible light polarization images are begin used in the fields of civil economic construction.The visible light polarization imaging technology is a novel optical reconnaissance method,which is able to obtain the intensity and polarization information of the object simultaneously.The performance of ground target detection and recognition can be improved significantly by fusing the two information.Meanwhile,the technology can increase the working distance of the imaging system,break through the limitations of the actual environment and use conditions,which has a wide range of application potential.However,the visible light polarization image is easily affected by external environmental factors.Such as atmospheric haze,uneven sun light,and small difference between the background and the target.All of this lead to the deterioration of the image quality and low definition of outdoor shooting,which affects the detection of the target,and leads to the huge reduction of the accuracy of automatic target detection.For target detection in visible light polarization smog images,it has become a crucial and unsolved problem that how to accurately detect ground targets.However,in the research of the visible light polarization smog images,the current smog removal technology usually ignores the distant scene recovery and texture information,whose training speed is slow and difficult.At the same time,the restored visible light polarization clear image has poor visual effects and cannot complete subsequent application tasks,such as target detection.In response to the above problems,our article has carried out in-depth theoretical research and experimental proof on the processing technology of removing smog from the visible light polarization smog image of the ground background.This thesis has main research contents and innovations,as follows:1)In the visible light polarization smog image removal smog field,we research a cycle convolutional neural network of the visible light polarization smog image removal algorithm.It focuses on the solution of the problem that artificial targets in the visible light polarization smog images are blocked by smog and the image quality is impaired.Based on the polarization characteristics of the visible light polarization image,the network generates a removal smog model in an end-to-end manner.The model calculates the polarization information of the visible light polarization image through the Stokes parameter.The target detection sub-network is used to detect the target smog area based on the polarization information,and then feature extraction network with a feature converter structure is proposed to generate the target smog-free area.By fusing the generated smog-free target area with the original smog visible light polarization image according to the corresponding coordinates,which obtained a clear smog-free image.More importantly,in order to obtain the final clear image,the rough clear smog-free image is again regarded as the input data of the de-smogging network,which puts our model in a loop topology.Experiments show that this method has better performance compared with the currently popular methods.2)In order to generate the global visible light polarization smog-free image with better visual effective,a multi-scale fusion conditional generation adversarial network remove smog algorithm is researched.Firstly,we use 4 different polarization angles of visible light polarization smog images as the input of the network.Secondly,use the generator and the fusion network to generate the feature figures of the visible light polarization smog-free images.At the same time,the original input image is characterized to obtain the polarization image,as the input of the enhancement network.Input the polarization feature figure and the visible light polarization smog-free feature figure into the PTM network.The final high quality visible light polarization smog-free image can be generated by comparing the output smog-free image with the real smog-free image are input the discriminator.Extensive experiments on the large scale real dataset demonstrate that the detail texture of the generated image by the proposed method is more clear than existing techniques.3)We constructed a large-scale dataset,which can be used for the evaluation of the visible light polarization smog images de-smogging.It contains 17,216 visible light polarization smog images and their corresponding real smog-free images.All images were captured by the visible light polarization camera,which collects different smoke images in different scenes and time periods.Different polarization information datasets can be derived based on this dataset.