Research And Implementation Of Face Super-resolution Reconstruction In Foggy Environment

Super-resolution reconstruction of face image is one of the key techniques to improve the resolution of face images and enhance visual effects.It has been widely studied and paid attention in the fields of computer vision,video surveillance and public safety.However,in extreme weather such as smog or when collecting face images from a long distance,it is affected by degradation factors such as illumination and noise,resulting in images with blurred,distortion,low resolution and other degradation defects,thus affecting the image or the actual application performance of the video.This paper analyzes the commonly used image defogging algorithm and super-resolution reconstruction algorithm by studying the related literatures of image defogging and super-resolution reconstruction technology at home and abroad.Aiming at the problem of unclear face image and low resolution in foggy video surveillance,defogging algorithm based on a scene depth and dark channel and a single image super resolution model based on auto-encoder are proposed,and designed a face super-resolution reconstruction system in a foggy environment.The main contents are as follows:(1)Firstly,in the defogging algorithm based on physical model,selecting different prior knowledge will directly affect the accuracy of estimating unknown parameters,which is very important for the final effect of image defogging.In order to adapt to different foggy environments(Such as dark light,different fog colors,etc.),this paper proposes a method based on scene depth and dark channel prior defogging,which firstly estimates the more accurate atmospheric light intensity by using the color change in the scene depth information,and calculates the observation intensity and estimated intensity of the scene atmospheric light.Absolute differences between the estimates of more accurate Transmission rate,and finally restore the image from the physical model and use the color balance algorithm to correct the visibility and color cast of the restored image.Compared with the traditional algorithm,the reference algorithm evaluation index the mean value of PSNR and SSIM are higher than other algorithms,and the mean value of PSNR and SSIM are increased by 1.6928 dB and 0.012 compared with the highest group.The mean value of the no-reference evaluation index NIQE is lower than other algorithms,which is decreased by 0.20107 compared with the lowest group.The experimental results show that the scene depth and dark channel prior defogging algorithm have better performance and clarity for foggy image restoration with different illumination and different tones.(2)Secondly,in the single-image super-resolution traditional model based on deep learning,the low-resolution images obtained by different degradation methods are experimentally analyzed and trained on the VDSR model.The bicubic method has a better model than the bilinear method.Training and performance indicate that different degradation methods affect the performance of traditional single-image super-resolution models.In this paper,with reference to the convolutional network layer in the VDSR model,a auto-encoder model that can simultaneously learn down-sampling and up-sampling is designed.The experimental results combined with the FERET face database show that compared with the traditional method,the PSNR value of the auto-encoder model reconstruction quality evaluation index designed in this paper is higher than other methods,which is about 0.59 dB higher than the highest value.Super-resolution reconstruction has better performance.(3)Finally,in order to verify the feasibility of the proposed algorithm in practical application,this paper designs a face super-resolution reconstruction system and a visual operation interface in a foggy environment.The main functional modules include: video reading,video frame extraction,face Image defogging and super resolution reconstruction.Experiments were carried out in the designed face reconstruction system based on the recorded foggy video dataset.The experimental results preliminarily proved that the proposed algorithm has practical application.