Study Of Hunger Behavior Of The Opleganthus Based On Video Computing

Being delicious and rich in protein,There is great economic value of breeding Oplegnathus.The regular and quantitative method was adopted in the traditional aquaculture industry.It is not possible to feed the fishes with the bait in time according to the starvation status.The problems such as untimely feeding or insufficient bait may easily occur,so the fishes of Oplegnathus often become starved,which has seriously affected the growing of the fish of Oplegnathus that brings economic benefits.The video calculation method is adopted to study the starvation behavior of Oplegnathus so that the starvation status of the Oplegnathus can be identified,which is of great significance to the cultivation of the Oplegnathus.In thispaper,we used deep learning technology to realize the target detection and velocity calculation of the Oplegnathus.In the end,LSTM is used to classify the starvation behavior of Oplegnathus speed series.The main work is as follows:(1)A detection method for the Oplegnathus based on multi-scale features is proposed.Under the experimental conditions,the image datasets of the Oplegnathus were constructed from above and below water angles including the surface image data set(AWOP data set)and underwater image data set(UWOP data set).As to the occlusion and blurred image,the lightweight neural network,MobileNet-SSD,was improved,and atrous convolution and SE block were added to the MobileNet network;in the overwater video and underwater video,the fishes of Oplegnathus may easily block each other,and to solve the problem,SE blocks were added to the feature maps of different scales to detect mutually occluded targets;Focal Loss function was used to calculate the classification loss,and to balance the proportion of background samples and target samples.The average detection precision of Oplegnathus by the method proposed in this paper respectively reached 91.87% and 87.45% in the overwater and underwater Oplegnathus datasets.High detection accuracy is achieved.(2)A velocity calculation method for the Oplegnathus based on monocular video is proposed.The algorithm in this paper includes three parts: detection,tracking and speed calculation.First,the modified MobileNet-ssd algorithm featuring good running speed and accuracy is adopted to detect the Oplegnathus.Then,the appearance featuretraining model of Deep sort was improved based on the appearance characteristics of the Oplegnathus.Through camera calibration we obtain the homography matrix of the image that is consistent with the real world,and map the pixels in the image to the coordinates in the real world using the mtrix;Finally,velocity calculation is done for the multiple fishes of the Oplegnathus in the video according to the frame time of the video.(3)A classification method for the hungry behavior of the Oplegnathus was proposed.Speed and shoal position are important characteristics of swimming fish.The position and swimming speed of Oplegnathus can reflect the biological status of fish.LSTM can selectively save and ignore the information in the unique gating mechanism,and it performs well in sequence learning.In this paper,the characteristics of the speed and position of the Oplegnathus were extracted,and then the sequence of the hunger characteristics of the Oplegnathus in different starvation states was constructed.The swimming speeds of the Oplegnathus were related to the positions of the fishes,and the hunger behavior classification of the Oplegnathus were realized by using LSTM algorithm.