Research On Benthic Image Recognition Technology Based On Deep Learning

In recent years,the total demand for benthos such as sea urchins,sea cucumbers,sea stars,and other benthos in the world has increased.The benthos aquaculture industry has achieved rapid development and created huge market benefits.However,many core problems have also been exposed,which are reflected in the fact that the most traditional manual fishing mode is still used,and there are core problems such as low efficiency and high risk factors.Therefore,it has become the main trend of the development of the benthic aquaculture industry to realize intelligent ocean and intelligent aquaculture utilizing the combination of benthic fishing operations and artificial intelligence.In recent years,artificial intelligence and machine vision technologies have achieved rapid development.These intelligent technologies have been widely used in the aquaculture field and can realize the function of real-time monitoring of the ecological environment.Combining the image recognition method in deep learning and deep mining the data in the process of aquaculture can improve the efficiency and decision-making reliability in the process of benthic aquaculture.The use of underwater robots to complete intelligent fishing is the key to development.The key technology for intelligent fishing is the intelligent identification of benthic organisms.In recent years,in the field of target recognition,the theory of deep learning has been continuously developed and gradually become an important tool for target recognition in actual production,and the technology of deep learning target recognition has made certain achievements in traditional recognition applications.However,the shallow water environment is complex,the color of benthos and the environment are poorly distinguishable,and benthos often presents a semi-shaded state,resulting in the low accuracy of target recognition using underwater robots.To make up for the shortcomings of the above benthos identification,this subject has researched underwater benthos target recognition.The main research progress is as follows:The data on this subject is from the benthos video in Xieton Town,Dalian City,Liaoning Province.Through the frame division of the video,4302 original pictures were obtained,and the data volume was increased to 13453 by using data expansion.Firstly,the image enhancement of underwater benthos is studied.The Contrast Limited Adaptive histogram equalization,the Multi-Scale Retinex with Color Restoration,and the Dark Channel Prior algorithm are compared.The quality of the enhanced image is evaluated,and the multiscale Retinex method with color recovery is determined as the enhancement algorithm to optimize the image,laying a foundation for the production of subsequent benthic data sets.Using the Labeling data set annotation tool,the underwater benthos image recognition data set,including holothurian,urchin,and starfish,is established in the form of rectangular annotation,with a total of 13453images.After image preprocessing,the benthos data set is produced.Secondly,the experimental analysis of the deep learning target detection algorithm is carried out.Analyze and compare representative target detection algorithms,such as SSD,YOLOv5,and Faster R-CNN.The experimental results were analyzed objectively and subjectively.According to the experimental results,compared with other algorithms,the SSD algorithm can ensure the recognition accuracy of benthos to a certain extent and is more suitable for improvement as the basic algorithm for subsequent benthos target recognition.This paper proposes a benthos identification algorithm based on an improved SSD target detection network.The method includes the following steps:firstly,the Receptive Field Block is used to expand the perceptual field of the low-level feature map,and the expansion convolution module is applied to downsample the low-level feature map to increase the weight of the low-level feature map on the final result map.Secondly,add attention mechanism to strengthen different depth features,give different weights to different original features,and calculate the final feature map,so that the results contain representative features,but also suppress irrelevant features,and finally fuse each feature map.The experimental environment for this project is:Intel（R）Core（TM）i7-9750H for the CPU,NVIDIA Ge Force RTX 2060 for the GPU,Ubuntu 20.04,keras2.1.5.To prove the recognition effect of the improved target detection algorithm,P-R curve,m AP₅₀,image visual effect,and other evaluation indicators are used to measure.The experimental results show that the improved SSD algorithm improves the m AP₅₀by 4.1%on the self-made benthic data set,effectively improves the detection accuracy of benthos,and reduces the missed detection rate and false detection rate of benthos.The improved SSD model is more suitable for the identification of benthos.In this project,the target detection algorithm is investigated along with some research on target detection algorithms in engineering applications.To facilitate the target detection function for benthic pictures,a visualization software for benthic target detection is designed and implemented.In summary,this project uses image pre-processing and image broadening to establish a benthic organism dataset,and the improved target detection algorithm for benthic organism target detection,which provides a new reference for intelligent fishing of underwater robots.