Research On UVMS Underwater Image Enhancement And Target Detection Methods

Underwater Vehicle Manipulator System(UVMS)is an important tool for marine resource exploration,and is widely used in offshore oil engineering,underwater salvage and other industries at home and abroad.At present,in the complex underwater environment,the autonomous operation of UVMS has always been a research hotspot.This subject is funded by the National Natural Science Foundation of China(51809128)and Zhenjiang City's Industrial Foresight and Generic Technology Project(GY2018018)to carry out underwater image enhancement and target detection research on UVMS.The main contents of the research are as follows:Firstly,UVMS program design and selection.UVMS can be roughly divided into sensing devices,execution devices,and control devices according to their functions.The UVMS hardware system mainly includes underwater robots,underwater manipulators,ground consoles,and power supply communication cables.The UVMS software system is mainly the main control system,perception system and host computer system.After separately designing the UVMS hardware and software systems,the underwater image enhancement and target detection methods are studied.Secondly,the fusion image enhancement method of UVMS based on multi-scale Retinex(MSR)optimization algorithm is studied.Firstly,the initial underwater image is filtered bilaterally to get the rough image and detail image.Then,the MSR optimization algorithm is used to process the rough image.Finally,the enhanced rough image and detail image are fused to get the output image.The fusion image enhancement method of UVMS based Retinex algorithm can effectively solve the problems of short effective sensing distance,low contrast,uneven illumination,and noise interference of UVMS under the condition of natural or artificial light source,which makes preparation for better underwater target detection of UVMS.Thirdly,underwater image data set is constructed and UVMS underwater image target detection algorithm based on RetinaNet is studied.Firstly,aiming at the problem that the number of underwater images is small,the classical expansion method is used to expand the data.Secondly,reduced the weight of the network and the amount of input data,DenseNet is used to replace ResNet to build the backbone network to improve the RetinaNet detection algorithm.The improved RetinaNet detection algorithm improves the accuracy and rapidity of underwater detection,it provides strong support for underwater visual servo control to capture sea cucumber.Fourthly,UVMS underwater experimental research.Firstly,the collected sea cucumber images are used to establish the image learning library,and then the data is expanded to enrich the data set;UVMS underwater experiment is carried out to enhance the UVMS underwater image,the fusion image algorithm based on Retinex,histogram equalization,multi-scale Retinex algorithm and bilateral filtering algorithm are used for image enhancement and contrast experiments;then the underwater target detection experiment is carried out,and the improved RetinaNet target detection algorithm is compared with RetinaNet algorithm and Faster R-CNN.Finally,the experiment of grasping sea cucumber with UVMS monocular visual servo control is carried out.Experimental results show that the proposed UVMS underwater image enhancement algorithm and target detection algorithm are effective and meet the expected requirements.