Research On Underwater Image Enhancement And Mosaic Method Based On Convolutional Neural Network

With the development of science and technology and the increasing demand for resources,the exploitation of marine resources and underwater exploration have become important development strategies of all countries.Due to the limited environmental conditions,including low underwater visibility,complex the terrain and strong deep water pressure,and Autonomous Underwater Vehicle(AUV)is required to perform underwater tasks effectively.AUV underwater detection operation is not affected by umbilical cord cable,and can conduct long time and wide range of underwater detection operation with high efficiency,high controllability,safety intelligence,and other advantages.However,the underwater imaging environment is complex and the image quality obtained by the AUV underwater camera is seriously degraded when compared to land images,and the single underwater image has a limited range of view,which makes it difficult to obtain sufficient underwater information Thus the visual range and resolution of underwater visual images are greatly limited.This thesis applies the convolutional neural network to underwater image enhancement and image registration,aiming to improve the quality of underwater images and obtain high-quality and wide-angle underwater panoramic images.The specific contents are as followsFirstly,underwater image enhancement method is researched on the basis of the trait of underwater image.The clear images in VOC2012 dataset are used to simulate underwater images through image blurring and color attenuation,and then one kind of convolution neural network framework composed of convolution subnet and deconvolution subnet is given,and the input and output of the network framework are full image.The nonlinear mapping from fuzzy degraded image to sharp enhanced image is formed by the hidden layer,and the convolutional layer is applied to study the fuzzy degradation characteristics in underwater degraded images,and symmetric of deconvolution layer effectively refines the convolution feature mapping details in order to gain underwater image more details information,so underwater image enhancement is realized in the endSecondly,underwater image feature extraction and matching method is researched.a improved CNN-RANSAC characteristics of underwater image registration method is put forward,and the improved VGGNet-16 network is applied to extract underwater image feature and generate robust multi-scale feature descriptor and feature points.After the feature coarse matching and the dynamic interior point selection,the improved RANSAC algorithm is given to eliminate false matching points in order to improve the robustness of feature point matching ultimately.Experiments of feature extraction and feature matching are carried out on a large number of underwater image datasets and the effectiveness of the proposed method is verified when compared with traditional SIFT and SURF registration algorithms.Finally,the method of image fusion and underwater panoramic image mosaic is studied.The frequently-used underwater image fusion methods are compared and analyzed,including direct average method,weighted average method,optimal seam fusion method and Laplace pyramid image fusion method,and the Laplace pyramid image fusion method adopted in this paper is analyzed in detail.Then the underwater panoramic image mosaic method is studied,which combines image enhancement,image registration and image fusion methods to obtain a panoramic underwater image with a wide view angle,high quality and no seam.The tunnel wall images,seabed images and lake bottom images collected by AUV are tested and analyzed by the proposed method.