Research On Underwater Image Enhancement Algorithm Based On Convolutional Neural Network

The ocean contains a lot of rich resources,and the exploration of marine resources currently relies on underwater images captured by underwater vehicles,but the underwater environment is more complex and uncertain than the terrestrial environment,and the underwater images are affected by light absorption and scattering,and there are problems such as color distortion and foggy blurring,which seriously limit the usable information from underwater images.In this thesis,underwater image enhancement techniques are investigated based on the principle of underwater image imaging,and the main research work is as follows.(1)In response to the problems of bias and loss of detail edge blur in previous algorithms for color correction,underwater image enhancement is divided into migration to underwater image de-colorization and overwater image defogging and deblurring based on the principle of underwater image imaging combined with convolutional neural networks.Migrating the complex underwater problem to a land-based environment with more mature applications allows for better implementation of underwater image enhancement.The experiments also verify the effectiveness of the method proposed in this study to divide underwater image enhancement into two stages of step-by-step processing for color bias correction and detail feature recovery.(2)Using the principle of underwater image imaging,the normalized residual energy ratio of different water types is adjusted to attenuate the effect of light scattering and highlight the absorption effect underwater to simulate the dataset with only blue-green bias for different water depths in two types of offshore and distant waters.This thesis also add a non-local attention module at the shallow level of the residual channel attention network to capture the long-term dependence between pixels,suppress noise and extract color style features.The residual network and the RCAN network with added non-local attention are successively trained on the dataset with only color bias.The experiments also validate the practicality of the simulated dataset and the effectiveness of the proposed network for color reduction.(3)In the defogging and deblurring phase,underwater image enhancement is migrated to the defogging and deblurring of aerial images,while a multi-stage progressive image restoration network is applied to the defogging and deblurring of images to train a new foggy sky dataset and adjust the network output.The network is divided into three stages,which include the first two stages based on U-net learning extensive pre and posttext information and the final stage using the original resolution sub-network to preserve image details and solve the image edge detail loss and blurring problem.The experiments also demonstrate that underwater images,after the color bias correction,image defogging and deblurring methods proposed in this thesis,achieve relatively excellent results in terms of qualitative and quantitative image quality evaluation criteria as well as the number of key points detected and identified compared with other methods.