Color Restoration And Enhancement Of Sand Dust Image Based On Deep Learning

Under sand dust weather conditions,due to the absorption and scattering of incident light by suspended particles in the atmosphere,the images collected by outdoor computer vision systems usually with yellowish color distortion,reduced contrast,and loss of detailed information.Such problems seriously affect the performance of outdoor computer vision systems such as video surveillance,video navigation and intelligent transportation.In this thesis,researches on the above issues are carried out,and the main work and the achievements are as follows:(1)Aiming at the problem that it is difficult to obtain a pair of sand dust image and their corresponding clear images as deep learning training samples,a sand dust image synthesis method based on a physical imaging model is proposed.Based on the imaging principle of sand dust images,that is,dust particles will absorb and scatter light to attenuate,and the attenuation degree of light of different colors is different.According to this characteristic,this thesis selects 15 color numbers that are close to the colors of sand dust images.The sand dust images under 15 different conditions are simulated,and finally constructs a large-scale dataset with a clear image and a pair of sand dust images.The experimental results show that the sand dust image synthesized by this method can train and evaluate the deep learning network.Related work has been submitted to the "Journal of Image and Graphics " and has been accepted.(2)Since traditional sand dust image enhancement methods do not perform well in color restoration,and the enhanced image still has problems such as blurring of details,a color restoration and enhancement method for sand dust images based on convolutional neural networks is proposed.First of all,to solve the problem of color distortion of sand dust images,based on the gray world theory,an image color restoration subnet is constructed to restore and correct the color of the sand dust images.Then,for details blurring and other problems,the adaptive instance normalized residual block is used to construct a de-dusting enhancement subnet,and the result image of the color restoration subnet is used as a condition to input the normalized residual block by the adaptive instance.Enhance the sand dust images.In addition,the perceptual loss is added to the network model to further constrain the details and semantic information of the restored image.The performance of the proposed method is verified by synthetic images and real images.The experimental results show that compared with the existing methods,this method obtains the highest average PSNR and average SSIM on the synthetic images,which are 18.7057 and 0.6695,respectively.Related work has been submitted to the " Journal of Image and Graphics"and has been accepted.(3)Since the existing dehazing methods based on convolutional neural networks can only ensure that the output image is very similar to the label image at the pixel level,it is easy to lose the detailed information of the image,and the phenomenon of excessive smoothness occurs,which affects the clarity of the image after dehazing.Based on this,a sand dust image color restoration and enhancement method based on generative adversarial network is proposed.Its basic framework consists of two generators and one discriminator.First,the first generator is used to restore the color of the input sand dust images,and then the color restored image is used as the input image of the second generator network,and the second generator is used to remove the sand dust in the image.In the discriminator,the block-based discriminator is used to learn the structural loss,rather than learning the entire image-level or pixel-level loss,so that the generated clear image is structurally closer to the real image.The experimental results prove that by adding a generative adversarial network,the enhanced result of the real sand dust image can be closer to the natural and clear image in terms of color,contrast,and brightness.