Research On Underwater Image Enhancement Algorithms Based On Convolutional Neural Network

Since the birth of mankind,human’s exploration of the ocean has never stopped,and the ocean not only provides us with the fishery resources necessary for survival,but also contains various mineral resources indispensable for social development.Especially in today’s environment of depleting exploitable resources,marine development has become a top priority for governments,however,the actual underwater images often have problems such as color distortion,blurred details,and insufficient contrast,which brings a lot of trouble to the subsequent scientific research work,so the research of underwater image enhancement technology is particularly important.At present,underwater image enhancement research is divided into two major categories,one is relying on traditional image processing technology to enhance images,and the other is underwater image enhancement based on deep learning,and its development has opened the door to a new world of image processing.The algorithms proposed in this thesis are based on the deep learning convolutional neural networks,and improve and innovate the networks.The experiment is carried out on multiple datasets,and the subjective visual effects and objective index scores are compared with various algorithms.The main work of this thesis is as follows:(1)Aiming at the problem of color distortion and insufficient sharpness of underwater images,an algorithm that integrates dual color space with multiple lightweight convolutional neural networks is proposed.Most of the classical depth learning algorithms only train in the RGB color space,which can result in some features being ignored and insufficient ability to correct color deviation,affecting the generalization ability of the entire network.The parallel operation of dual color space can make up for this defect.The HSV color space is introduced while training on RGB color space,and combined with residual network and attention mechanism to achieve the organic integration of dual color space and multiple subnetworks,which makes up for the problems of partial color information loss and incomplete feature extraction in the training process of singlecolor space,and improves the generalization ability of the entire algorithm while ensuring the enhancement effect.Experimental results show that the proposed algorithm can not only correct the color deviation of underwater images,improve brightness,enhance the subjective visual effect of images,but also improve the objective index score of images.(2)Aiming at the problem that existing deep learning algorithms have certain limitations in carrying out enhancement tasks,an algorithm combining physical imaging models and convolutional neural networks is proposed.General deep learning enhancement algorithms rely on a large amount of training data and are difficult to quickly and accurately find the connections between various features,increasing training costs.The introduction of physical imaging models can solve this problem,itcan enrich the single end-to-end learning method of deep learning,organically combine the abstract features extracted from the network with its actual physical significance,and provide a solid theoretical support for the optimization of the training model.The underwater imaging model has two important parameters: the direct transmission parameter and the background scattering parameter.In this thesis,the direct transmission parameters are further divided into three levels of transmission corresponding to three different hierarchical subnetworks,and the direct transmission is also divided into shallow feature direct transmission and deep feature direct transmission.Similarly,there are corresponding subnetworks.The parameters of the imaging model correspond to the subnets with different functions one by one,which greatly enhances the generalization ability of the model as well as enhances the enhancement effect.Experimental results show that compared with other enhancement algorithms,the proposed algorithm can effectively enhance the edge and texture details of complex underwater images and improve image clarity.