Research On Underwater Target Detection Method Based On Deep Learning

In recent years,with the boom of the seafood market,traditional manual fishing methods can no longer meet the market demand.Therefore,using robots to realize automatic fishing instead of manual operations has broad development prospects.As a typical seabed creature,sea cucumber has always been a key fishing target,and the research on target detection of underwater marine creatures has become a key field of marine development.This paper has conducted an in-depth analysis of the research status of related fields at home and abroad,taking sea cucumbers as the research object,mainly studying the methods of the detection,counting and positioning of sea cucumbers.In a complex ocean environment,in view of the poor detection accuracy of traditional underwater target detection methods and the low efficiency of three-dimensional positioning,an underwater target detection method based on deep learning is proposed.This paper first enhances the underwater image,then classifies the target based on the deep learning method,and then counts it,finally,the detection frame is taken as the target area of interest and combined with image segmentation to obtain the capture position of the sea cucumber and convert it into three-dimensional coordinates to achieve stereo positioning.The main research work is as follows:Firstly,a sea cucumber underwater data set is established.In view of the small scale of the underwater sea cucumber data set,the method of data set enhancement is used to expand the data set to provide a data basis for convolutional neural network training;Aiming at the problem of underwater image quality caused by the absorption and scattering of underwater light,an underwater image enhancement processing method based on white balance and image fusion based on color compensation is improved.Secondly,the target detection algorithm is studied.This paper analyzes the mainstream target detection algorithms,and in view of the actual situation of underwater detection of sea cucumbers,the mainstream detection network has the problems of slow speed and target missed detection problems.The target detection algorithm is researched and improved from four aspects: feature extraction network,convolution structure,multi-scale fusion method,and up-sampling method in this paper,so that the sea cucumber target can be detected quickly and accurately.Thirdly,in order to obtain the distribution of underwater sea cucumbers and measure the mining significance of the fishing area,the method of counting underwater video sea cucumbers is studied.In view of the current target counting method,the counted targets will be counted repeatedly in subsequent video frames,an underwater target counting method based on target library comparison is proposed,which compares the target that appears again after disappearing from the video frame with the comparison library,which effectively avoids the problem of repeated counting of this type of target,and can accurately count sea cucumbers.Finally,the positioning method of underwater sea cucumbers is studied.Aiming at the problem of decreased matching accuracy caused by poor underwater image quality,an underwater image enhancement methods is used to increase the number of matching points,and an improved GMS method is proposed to filter out incorrect matching points for the higher rate of false matching of underwater sea cucumber features to improve matching accuracy;In view of the long time for traditional binocular stereo matching to perform binocular stereo matching on the entire image,this paper reduces the range of feature search and improves the speed of target positioning by binocular matching on the region of interest.In addition,in order to obtain more accurate capture point location information of sea cucumbers,a method combining image segmentation to perform ellipse fitting on the region of interest is used to obtain the sea cucumber center point position and perform coordinate conversion on it to achieve underwater stereo positioning of sea cucumber targets.