Research On Underwater Image Quality Enhancement And Object Detection Algorithm

The ocean contains abundant resources.The development and utilization of marine resources have attracted the attention of all countries.Effective perception of the marine environment is a prerequisite for the development and utilization of marine resources.Underwater image processing based on computer vision technology is an effective means to perceive the marine environment.However,the absorption and scattering effects of light in water will lead to degradation of underwater images.The image quality enhancement method can effectively solve the image degradation problems such as color deviation,haze and contrast reduction of underwater images,and is an important way to improve the quality of underwater images.In addition,in addition to the problem of degraded underwater images,the characteristics of underwater biological objects,such as dense,small scale and similar color to the background,bring challenges to underwater object detection based on computer vision.In this paper,the underwater image quality enhancement algorithms with different application requirements are designed for degraded underwater images.And a marine biological object detection scheme is proposed to help human eyes perceive the marine environment.The details are as follows:(1)An image quality enhancement method based on underwater physical imaging model is proposed to solve the problem of underwater image degradation.First,homomorphic filtering is used to remove the color cast problem in underwater images.Then,in view of the problem that the image restoration algorithm based on a single transmission image cannot adapt to different underwater environments,a double transmission map dehazing algorithm is designed to adapt to different underwater environments and improve the clarity of the image;further,the color line model is used to estimate the transmittance.The quasi-norm relaxation optimization model and the improved proximal alternating linear minimization method are used to solve the problem,which improves the accuracy of transmittance estimation,so as to obtain a clearer underwater image.(2)Aiming at the problem of serious hazing and poor image quality in underwater video images,a generative adversarial network based on color line loss function is proposed to achieve underwater image quality enhancement.First,in order to ensure the training quality of the generative adversarial network parameters,a comprehensive training dataset is proposed to train the neural network.Then,based on the principle that the color line of the fog-free image passes through the origin of the RGB space,and the color line of the foggy image deviates from the origin of the RGB space,the loss function of the color line is designed to improve the dehazing ability of the network.Furthermore,the discriminator is designed as a dual-discriminator structure including a color line loss branch and an adversarial loss branch to guide the generator to achieve image dehazing and content preservation at the same time,and promote the generator to generate clear underwater images.Experimental results show that the proposed method improves the quality of the generated images while ensuring the processing speed.(3)An object detection algorithm with multi-attention path aggregation network is proposed to solve the problem of low detection accuracy caused by the characteristics of dense and small size of marine biological objects.First,aiming at the problem of large target scale distribution spans,a path aggregation network structure is designed to deal with multi-scale problems.This structure introduces the different scale features of the backbone network into a bottom-up path,which enhances the semantic features and improves the feature extraction ability.Then,a multi-attention mechanism combining coordinating competitive attention and spatial supplementary attention is proposed to weight the detected objects on the feature map,which improves the detection accuracy of small objects and dense objects of marine organisms.Finally,the underwater images are enhanced using the image quality enhancement algorithm,which improves the visibility of the detected images.(4)Aiming at the problem that the object detection recall rate will be reduced after the quality of the original underwater image is enhanced,resulting in the reduction of detection accuracy,a generative adversarial neural network framework for learning the distribution of the original data is proposed to learn the enhanced data and the original data at the same time.The detection accuracy is improved on the basis of image quality enhancement.First,the proposed network framework inputs the enhanced image and the original image into the same discriminator of the generative adversarial network model,enabling the generated enhanced images to learn the original data distribution.Then,the proof of JS divergence preservation property and convergence for original data and enhanced data proves that the proposed framework can learn the original data distribution.Finally,experiments verify that training the object detection network with the enhanced image datasets generated by the proposed method can ensure high detection accuracy.