| ROV is an important underwater tool,which is mainly used in marine resource exploration,fishery,hull maintenance and pipeline detection.Visual system is an essential part of ROV and plays a vital role in underwater operation.At present,many underwater projects still need divers to work,but the real water environment is complex and changeable,so it is very important to use underwater ROV instead of manual work.Under the support of the National Natural Science Foundation of China(NSFC)project(5180912)and the Changzhou International Science and Technology Cooperation Project(CZ20210013),we studied the ROV underwater target detection method:First,the overall system design of ROV.Hardware includes all kinds of sensors,drivers,ROV underwater camera,floating materials and water control box,etc..The software includes underwater control system,communication module,Data Acquisition Module,training set of underwater small target based on depth learning,underwater image enhancement network environment and underwater target detection network environment.Finally,a ROV system suitable for complex water environment is designed.Second,the research of ROV underwater image enhancement is compared with several traditional underwater image enhancement methods,and a new ROV image enhancement method based on GAN network is proposed,an improved network structure based on GAN network is designed,including generator and discriminator.The feature extraction module FEM is proposed to extract the feature depth of ROV image,and the enhancement result of Leaky Relu activation function is selected,through the discriminator training to improve the image enhancement effect,the final completion of the ROV image enhancement.Thirdly,based on the background of ROV underwater target detection,the traditional one-stage detection method and two-stage detection method are analyzed and compared,an improved ROV target detection network architecture based on Faster R-CNN is proposed.In the structure of Main Body Network,RESNEST network is used to replace VGGNet network,and a module for feature extraction of small underwater target is proposed,which changes the original feature extraction method,finally,considering the edge overlap of the target detection frame,the improved loss function is used to optimize the whole network structure to improve the accuracy of the regression and improve the ROV target detection accuracy.Fourthly,the performance of the whole structure of ROV is tested,the data set is built according to the target detection network,and the ROV image enhancement method is verified by enhancing the collected underwater images,after the whole ROV system is designed and assembled,the ROV underwater target detection experiment is carried out.The experimental contents include ROV performance test,Underwater Image Enhancement Test,and real water environment target detection experiment,which verifies the proposed target detection method. |
