Application Of Convolutional Neural Network In Underwater Image Intelligent Processing And Recognition

Human exploration of underwater space has been an emerging activity in the past decades with the continuous progress of science and technology in this domain.Underwater optical imaging is one of the most significant techniques for comprehensive underwater environment-related situational awareness and it is highly demanded in various engineering practices,such as underwater resource exploration,underwater rescue,underwater archaeology investigation,underwater defense activities and so on.However,the imaging capability in distance of an ordinary imaging system is considerably jeopardized due to the scattering and selective absorption of light by the water body and suspended particles.Additionally,foggy,blurred,color-distorted and other effects could be easily identified for most of the underwater images captured by the ordinary imaging systems,the image quality could be even worse in a turbid water environment.Unlike the ordinary imaging systems,laser-assisted underwater gated imaging technology can effectively suppress backward scattering and enhance the distance of underwater imaging,but the images obtained by this technique are also coupled with problems such as high noise,low brightness and low contrast.In this thesis,the image enhancement algorithm research was carried out for the improvement of imaging quality of conventional underwater imaging and gated imaging.Based on the results of above-mentioned studies,intelligent detection of underwater objects has been achieved by establishing a imaging dataset of underwater specified objects and applying convolutional neural network for objects recognition and classification.The research results of this thesis are of great potential in promoting the development of underwater imaging technology,they might highlight a path to intelligent underwater optical imaging and detection.The main content of research of the thesis is as follows:Firstly,new ideas of algorithms for intelligent image processing and recognition based on two underwater camera image enhancement algorithms and convolutional neural networks were proposed through the investigation of underwater imaging technology,image characteristics as well as different methods for image enhancement and intelligent recognition.In this section,several classical underwater image enhancement algorithms were studied and reproduced while focusing on the respective advantages and problems of the different existing algorithms,in order to investigate the challenges from image quality degradation occurring in both conventional and gated images.To further improve the performance of the existing algorithms,a multi-scale fusion-based underwater image enhancement algorithm was proposed for conventional image enhancement while a non-local mean denoising algorithm based on increased gray-scale stretching and sharpening was proposed for underwater gated image enhancement.After subjective scoring of image enhancement effect and objective image quality evaluation including to evaluate the image quality before and after enhancement,compared with other algorithms,in both subjective and objective evaluation for image enhancement of two sets of underwater cameras.Being evaluated by both subjective scoring and objective assessment on image information entropy,contrast and other indicators,the algorithms in this thesis achieved better results in comparison with other algorithms investigated in this thesis.Last but not the least,in this study,two cameras were used to capture images of specific underwater targets to build the dataset.Then the network model was trained based on this dataset using the deep residual network Res Net34.The network model was used to realize the recognition and classification of specific underwater targets and improve the automation and intelligence of optical detection instruments.