Research On Sand Dust Image Enhancement And Restoration Methods

Currently,computer vision systems are widely used in video surveillance,intelligent driving,smart cities and other fields.However,the performance of these vision systems is directly related to the quality of the captured images.Under sand dust weather conditions,due to the absorption and scattering of sunlight by sand dust particles suspended in the atmosphere,the captured images have low contrast,blurred details,small dynamic range of images,color distortion and other degradation phenomena.If these feature corrupted images are directly used as input information of computer vision system,the performance of computer vision system for feature extraction,analysis and understanding will be in-evitably reduced.Therefore,through the research on the enhancement and restoration methods of the snad dust image,and how to effectively reconstruct sharp images from the sand dust images,which has important theoretical and practical significance for improv-ing the performance of the computer vision system and enhancing the practical application of the computer vision systems.In recent years,researchers have carried out a large number of enhancement studies on images in severe weather,such as haze,rain and snow,and achieved remarkable re-sults.In addition,the researches on image enhancement and restoration under severe sand dust weather conditons have also increased significantly.Through a comprehensive in-vestigation of the current states of domestic and international research in the field of haze image,sand dust image and low-light image enhancement and restoration,it is found that the current research methods for sand dust images still have the following problems:(1)The color casts correction methods for sand dust images are not well studied,which makes the recovered images still have color casts,and even introduce new color deviation.(2)Depending on various prior assumptions,once the scene conditions do not satisfied with these prior conditions,it is difficult to restore the ideal results.(3)It depends on the synthetic data to drive the depth learning model,but the synthetic data is difficult to characterize the real sand dust scenes,which will lead to domain shift of the deep learning model.It is difficult to obtain ideal results with models trained by using synthetic data to process real sand dust images.(4)It focuses on the processing of sand dust images with high brightness,and less attention is paid to images captured in low-illumination sand dust environments.Moreover,due to the worse and more complex imaging conditions,the existing sand dust image enhancement algorithms can not obtain good results when they are used to directly recover low illumination sand dust images.To solve the above problems,this dissertation combines the theoretical knowledge of atmospheric scattering model,human vision model,and zero-shot learning,and carry out enhancement and restoration studies on images captured under sand dust weather con-ditions from different perspectives.The research contents and contributions of this paper focus on the following aspects.(1)To solve the problems of new color casts introduced by the enhanced images,and the difficulty of accurately estimating atmospheric light and transmission.According to the attenuation characteristics of the color channel in the sand dust images,the red chan-nel is proposed as the reference to correct the color deviation.The atmospheric scattering model based on the blue channel is improved.The transmission is calculated by assuming that the maximum and minimum channels of the sand dust image are positively correlated with the transmission,and the atmospheric light is estimated by using the channel differ-ence.The proposed method solves the problem of correct estimation of atmospheric light and transmission,and finally obtains vivid colors.To solve the problem that it is difficult to balance the quality of the recovered image and the computational time,the red and green channels are used to estimate the rough transmission at the pixel level,and then the attenuation difference between the color channels is used to obtain the compensated transmission,thus ensuring the accuracy of the transmission estimation.The proposed method can better improve the algorithm speed and ensure the image restoration quality.(2)To avoid image enhancement and restoration quality were affected by the priors and synthetic data,a network for simultaneous estimation of atmospheric light,transmis-sivity and clear image is designed based on the theory of atmospheric scattering model and zero-shot learning method.Through joint learning network model,transmission and atmospheric light are estimated,and finally the enhancement and restoration of sand dust image are achieved.In addition,to address the problem that the enhancement of local dark regions in the recovered image is not obvious,and it is difficult to enhance the image details,a depth network model for estimating the brightness map is proposed based on Retinex theory,and thr degraded image is generated by perturbing the input image with a perturbation factor.Using the image pairs formed by the captured sand dust images and the generated degraded images,a zero-shot learning method is used to train the network model to obtain the brightness components.Finally,the sand dust images are enhanced by the Retinex model,thus solving the problem of uneven local and global enhancement in the images.(3)Because the low-illumination sand dust images are captured under low illumina-tion or nighttime dusty environment,the poor imaging environment makes low-illumination sand dust images have more complex characteristics,and it is more challenging to enhance the low-illumination sand dust images.Therefore,based on the study of low-illumination image and sand dust image enhancement algorithms,a multi-path color casts correction method is proposed.Finally,the image is enhanced,noise is reduced,and texture details in the image are highlighted by the way of transmission guidance.The method is effective in processing low-illumination sand dust images,which can eliminate color deviation and improve the brightness and contrast of the image.