Research On Underwater Image Enhancement Technology Driven By Deep Learnin

The ocean is an untapped treasure trove of riches.Nowadays,the population is rising rapidly,the environment is deteriorating,and the terrestrial environment can no longer carry the resource needs of human beings.The development of the oceans is already an extremely urgent strategic requirement.Since the ocean is both a carrier of information and an expression of the underwater scene,it plays an irreplaceable role in detecting and perceiving the underwater environment,therefore,underwater image enhancement remains a difficult task to be solved.However,the propagation distance of light underwater is affected by various reasons,resulting in greenish,bluish and white haze problems in underwater images.The contribution of deep learning in underwater image enhancement has partially solved the color bias problem,but the deep learning approach relies heavily on the quality of the dataset,and the collection of the original underwater image dataset is more difficult,which makes the progress of deep learning in underwater image enhancement task slow.The lack of access to paired highquality datasets has a knock-on effect on the convolutional neural network approach.The learning of the convolutional neural network approach is obtained exclusively from the feature maps extracted from the dataset,making it difficult for the convolutional neural network-based underwater image enhancement approach to solve the blurring and white mist problems and to correctly recover the high-frequency detail loss in some underwater images.With the advent of generative hostile networks,pursuing dataset quality is no longer the only solution,resulting in generative adversarial networks can learn the texture features of the image and generate the recovered image of underwater images autonomously.Due to the constraint of image content loss,the generator does not generate unconstrainedly,depending on the weights of various losses,limiting the generator to generate underwater images that are structurally consistent with the original image,while generating autonomously in high frequency details.Therefore,multiple loss functions are needed to cooperate with the constraints.In this paper,we discuss and summarize the advantages and disadvantages of the above underwater image enhancement methods and propose a more stable underwater image enhancement based on conditional generative adversarial network to enable the network to obtain more stable output and at the same time obtain good enhancement effect.The work accomplished in this paper consists of the following parts.(1)In this paper,we use the most important underwater physical image model for traditional underwater image recovery as model input for open-source RGB-D data,and use depth information to adapt to the depth information of the underwater image.(2)We use L1 and MSE loss to constrain the pixel level in content loss,structural similarity loss and feature loss together in structure and feature level respectively,and adversarial loss in local blur level to generate high frequency information autonomously in order to obtain clear underwater images.(3)In this paper,we also improve the generative adversarial network in two aspects.First,in feature learning,the combination of multiscale dense blocks with residual connections and front-and back-end attention mechanisms in the generator allows the features extracted from the front end of the generator to be fully utilized in the back end.The combination of multiscale dense blocks extracts the front-end features at multiple scales to obtain features at different scales,and the combination of channel attention mechanism and spatial attention mechanism allows the features extracted using multiscale dense blocks at the front-end to be filtered,removing the meaningless features and keeping the features that are important to learn.On the other hand,in terms of network robustness,the traditional 0-1 classification is abandoned and a more continuous value is used to discriminate the generated images of the generator.Throughout the network,a convolutional layer with a step size of 2 is used instead of pooling layers,which can use computational effort instead of brute force information discarding.(4)Finally,two non-reference evaluation metrics that are highly consistent with human senses are used to give a subjective and objective evaluation of the experimental results to prove the superiority of the proposed algorithm in underwater image enhancement.The proposed method achieves better results in terms of contrast,brightness and blurring of underwater images and the robustness of the network.The proposed method achieves better results in terms of contrast,brightness and blurring of underwater images and the robustness of the network.