| With the improvement of artificial intelligence technology,computer vision as its core has also entered a high-speed development stage.Obtaining high-quality basic images is the key prerequisite to ensure the effective performance of computer vision tasks.However,in the underwater environment,due to the absorption and scattering of light by suspended particles in the water medium,underwater images usually have problems such as color distortion and blurred details.Similarly,in foggy environment,the presence of tiny particles in the air can cause images blurred and low fidelity.Lowquality images in these two environments not only bring poor visual perception to humans,but also limit the performance of computer vision tasks such as target detection and semantic segmentation.Therefore,it is important and valuable to improve degraded images.In this paper,by studying and analyzing various image recovery methods,we investigate the degraded image recovery in underwater and foggy environment according to the image imaging characteristics of two environments,respectively.The details of the research are as follows:(1)To address the problem that the existing underwater image recovery methods have poor robustness and cannot handle complex underwater scenes and the recovery effect is not obvious,this paper proposes a underwater image recovery method based on Swin Transformer generative adversarial network.The method firstly decomposes the image by using discrete wavelet transform.Then the method introduces Swin Transformer with strong feature extraction ability to build a color reduction network to eliminate color distortion of low-frequency image,and introduces residual block to build a detail enhancement network to enhance the blurred details of high-frequency images.Finally,the network is trained with the designed hybrid loss function to obtain realistic and texture-rich underwater images.In addition,this paper proposes an underwater image synthesis method that can compensate for the lack of underwater data sets according to the imaging characteristics of underwater images.The experimental results show that the proposed method has good robustness and achieves better visual effects and objective performance on both real datasets as well as synthetic datasets.(2)To address the problems that the dark channel prior algorithm is prone to distortion when recovering the sky region and the lack of details in the defogged image,this paper proposes an improved dark channel defogging algorithm based on sky segmentation.First,the sky region is segmented from the fogged image by the proposed two-stage sky segmentation method.Secondly,the atmospheric light value is estimated based on a multi-scale window to improve the estimation accuracy and eliminate the limitation of a single window.Then the image is processed with a super-pixel window and the transmittance of the sky region is adjusted with the help of the bright channel to make the transmittance estimation of the whole image more accurate.Finally,the edge-enhanced image is used as a guide image and combined with the gradient domain guide filter to refine the transmittance to enhance the image details.The experimental results show that the proposed method solves the problem of distortion of the sky region in the defogged image,and the processed image is clearer and more detailed.(3)In this paper,degraded images and images processed by the proposed method are applied to different vision tasks and compared.In the underwater environment,edge detection and diver detection tasks are carried out.In the foggy environment,feature point detection and target detection tasks of driving scenes are carried out.The experimental results show that improving degraded images can largely improve the performance of subsequent related vision tasks. |
