Research On Image Acquisition And Restoration Method Of Underwater Vehicle

As an underwater operating equipment,underwater robot's recognition,detection,tracking and positioning of underwater targets have always been the focus of visual system research.The absorption and attenuation of light by water body and the impurities contained in it will reduce the quality of visual imaging of underwater robot,and the image information of underwater target cannot be clearly presented,which has become an important bottleneck restricting the development of intelligent underwater machine vision.With the development of deep learning,the visual system of underwater machines and equipment should be combined with deep learning to improve image quality and restore target image information so as to better complete underwater detection and recognition tasks.In order to improve the accuracy and reliability of underwater vehicle operation,the research on the visual acquisition system of underwater vehicle and the method of underwater visual image restoration was carried out.The specific research work was as followed:Aiming at the problems that traditional underwater image restoration methods rely on prior knowledge,generalization and poor robustness,a Double Branch Feature Extraction Network was designed based on the capability of complex nonlinear system modeling and feature information extraction of convolutional neural network.DBFEN and the Multi-Scale Attention Fusion Network(MSAGFEN)target two models for underwater image restoration.DBFEN model was based on the encode-decoding structure and constructs a double-branch feature extraction structure to solve the difficult problem of feature extraction.MSAGFN was an improved structure of DBFEN,which focused on strengthening the extraction and utilization of feature information.It combined multi-scale feature extraction with attention mechanism to extract effective information and suppress invalid information.The experimental results showed that both MSAGFN and DBFEN restoration models could effectively improve the visual effect of underwater images,and made the images clearer,color and contrast corrected,which was close to the images in natural scenes.It was difficult to obtain clear images corresponding to underwater images,which made deep learning network model unable to carry out effective data training.In order to solve the shortage of training data set,the natural scene images were mapped to the simulated underwater environment by generating adversarial neural network to generate three common underwater image data types.After testing,the synthetic underwater image can simulate the real underwater environment well,and also help to train the model fitting effect.In order to accomplish the task of visual image acquisition of underwater robot better,the selection of visual image acquisition system and underwater camera which was composed of main controller,image acquisition card,underwater camera and artificial light source is studied.Combined with the vision system,the paper studies the brushless motor speed regulation system,analyzed the principle and process of the brushless motor speed regulation controlled by the electronic governor,and used the PWM control signal generated by the main control board to realize the three-state control process of the positive and negative rotation and stop of the propeller motor.The corresponding software processing system was designed according to the studied AUV visual acquisition hardware system,which provided a process basis for the realization of AUV visual image acquisition and image enhancement and restoration.Figure[52]table[8]reference[90]...