Research On Quantitative Recognition Algorithm Of Marine Organisms Based On Deep Learning

Marine benthos is an important ecological group of marine organisms.It is an indispensable part of the marine ecosystem.Due to the complex environment and diverse biological composition of the seafloor,the detection and counting of marine benthos has been an important research direction in the field of marine biological object detection.The detection and analysis of marine benthic organisms can help researchers analyze the composition and evolution of ecological communities as well as the habitat and reproduction patterns of organisms.It is conducive to protecting the deep-sea ecological environment and promoting the sustainable development of marine ecology.In this paper,the feature extraction mechanism based on CNN and the feature extraction mechanism based on Transformer are analyzed,the advantages and disadvantages of single-stage object detection algorithm and two-stage object detection algorithm are compared.In addition,different quantitative detection schemes are proposed for static data and dynamic video data of marine benthic organisms,which lay the foundation for the realization of quantitative identification of marine benthic organisms.The quantitative detection algorithm YOLOT(You Only Look Once with Transformer)is proposed for the characteristics of marine benthic organisms.In addition,the multi-target tracking algorithm is studied and a single-hypothesis tracking and matching framework UD-Deep SORT which combines Kalman filter and Hungarian algorithm is proposed.Through experimental validation,the object detection model proposed in this paper improves the accuracy of marine benthic organism detection counting in images,while improves the accuracy of target matching in video.The main research work of this article are as follows:(1)To address the problems of color bias and blurring in images caused by factors such as suspended particles and small molecules in water,the similarities and differences between underwater imaging and atmospheric imaging models are investigated.The MSRCR(Multiscale Retinex with Color Restoration)and the Dehaze Net defogging model based on convolutional neural network are used to pre-process the experimental dataset to support the subsequent data annotation and experiments.(2)The quantitative detection algorithm YOLOT is proposed for marine benthic organisms in the static data.Firstly,in order to reduce the influence of complex seafloor environment on detection and improve the feature extraction ability of the backbone network,a feature extraction module based on attention mechanism is introduced.Secondly,a feature fusion structure based on Transformer is designed.Lastly,a positive and negative sample selection strategy based on Gaussian probability distribution is introduced to reduce the influence of the positive and negative sample imbalance on the training phase of the network and improve the training efficiency of the network.The experiments show that the detection accuracy of YOLOT in URPC dataset reached84.44%,which is 9.09% better than YOLOv4.Therefore,YOLOT can quantitatively detect marine benthic organisms accurately.(3)To address the problem of quantitative biological identification in dynamic video data of marine benthic organisms,the multi-target tracking algorithm is studied in the paper and the UD-Deep SORT object tracking algorithm based on Deep SORT is proposed.On the one hand,the algorithm introduces the unscented Kalman filter to improve the trajectory prediction accuracy of the algorithm.On the other hand,the algorithm uses DIOU(Distance-Intersection Over Union)to calculate the matching degree of the anchor frame,which optimizes the matching mechanism between the prediction frame and the detection frame.The experiments show that the tracking accuracy of UD-Deep SORT proposed in this paper is improved by 2.58% over Deep SORT,which is of high application value for studying the distribution of marine benthic organisms.