Research On Underwater Image Enhancement And Super Resolution Algorithm

As the population grows,the consumption of resources on the land is accelerating,and the ocean occupies most of the earth and is rich in biological and petroleum resources,so people are looking at the vast ocean for future development.my country attaches great importance to the development of marine resources.my country has also proposed a strategic policy for marine ranches,aiming to build a marine ecological ranching rich in food and sustainable development.In order to keep the marine environment from being destroyed as much as possible,many resource exploration work It is carried out by underwater robots,and underwater target segmentation and target recognition based on optical vision are the key to the intelligent work of underwater robots.However,the lack of color channels caused by the turbulence of the underwater environment and the suspended matter in the seawater causes the scattering of light,the defocusing of the camera,and the attenuation of the light.The underwater image presents the problems of color cast,lack of detail,and low contrast.The color distortion,low contrast,lack of detail and blur of underwater images have brought great challenges to the target recognition and target detection of underwater robots.In order to improve the quality of underwater images,in this thesis i proposes an underwater image enhancement and image super-resolution algorithm for the color cast,low contrast,lack of detail and blurring of underwater images.Image enhancement based on traditional algorithms is researchde to deal with the problem of color shift and low contrast of underwater images;then the image super-resolution algorithm based on convolutional neural network solves the problem of missing details and blurring of underwater images,thereby elevating the quality of the underwater images.The main contributions of this thesis are as follows:(1)Based on the Color balance and fusion for underwater image enhancement(UWB)underwater image enhancement method,the compensation method for the red and blue channels is improved,and the red and blue channels are adaptively adjusted through the standard deviation of each channel of each pixel.Compensation,and a gray-scale world algorithm based on white point detection is used to correct the color shift of the compensated image.At the same time,based on the dark channel prior defogging algorithm,the transmittance is estimated by superpixel segmentation instead of uniformly segmenting the local block,and in this process,an image adaptive weight factor is used to estimate the transmittance map to complete the image contrast elevate.(2)By analyzing the degradation process of underwater images,including the blur types of underwater images such as Gaussian,defocus and motion,and analyzing the limitations of existing degradation models,through Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels(DPSR)For the fuzzy kernel estimation idea of the unknown blur type in the image super-resolution network,an adaptive blur function estimation method based on DPSR is used for the estimation of the underwater image degradation model to include various blur processes in the underwater environment.(3)Based on the Residual Dense Network for Image Super-Resolution(RDN)network,in response to the problem of too much low-frequency information and too much parameter in the global fusion process of the dense residual network,the residual dense block(RDB)in the network Adding and series heterogeneous structures,and restricting the gradient value through the Smooth absolute error(smooth L1)loss function,so that it can be globally derivated during the calculation process and has higher robustness to outliers.Finally,by adding parallel sub-pixel convolutional layers to train different scale factors,the superresolution of the image is realized.Finally,after experimental comparison and analysis,the algorithm proposed in this thesis can effectively solve the problem of color shift and low contrast of underwater images,and can greatly improve the detailed information of underwater images and improve the visual effect of underwater images.Finally,a target detection method,YOLOv3 is used for experimental verification.The results show that this algorithm can greatly improve the accuracy of underwater target recognition,thereby verifying the effectiveness of the underwater image enhancement and super-resolution algorithms proposed in this thesis.And engineering applicability.