Research On Intelligent Identification And Counting Of Sea Foods In Marine Pasture Based On Deep Learning

The identification and counting of sea foods is an important link in the aquaculture process,which provides an important basis for seedling release,breeding status monitoring,precise feeding and economic benefit estimation.Aiming at the problems of high cost,low efficiency,and difficulty in guaranteeing the counting accuracy of sea foods identification and counting methods in the current breeding process,this paper takes the sea foods in the real natural bottom-sowing breeding environment as the research object,and uses the underwater video of sea foods as the data source.A systematic study has been carried out on the identification and counting of sea foods.The specific research contents are as follows:First of all,aiming at the problems of low efficiency and difficulty in guaranteeing counting accuracy in the traditional sea treasures identification and counting method,this paper designs a multi-category sea foods counting method based on video multi-target tracking.The YOLOv7 algorithm with excellent performance is used to realize the sea treasure target detector.Using the multi-target tracking idea of the BYTE algorithm,a multi-category trajectory generation strategy and a counting strategy based on the trajectory ID number are designed,and a multi-category sea treasure tracking and counting method is proposed.In addition,in order to evaluate the performance of the proposed method more objectively,a set of evaluation indicators that are more suitable for the method based on trajectory ID number counting is proposed.The experimental results show that in the improved average counting precision MACP(Modified Average Counting Precision),improved mean absolute error MMAE(Modified Mean Absolute Error),improved root mean square error MRMSE(Modified Root Mean Square Error)and frame rate FR(Frame Rate)In terms of performance,the methods in this paper are 91.62%,5.75,6.38 and 32 frames per second,and almost all indicators are better than other algorithms compared in this paper,especially the MACP and FR indicators are 29.51% and 28% higher than YOLOv5+Deep SORT,And the indicators of MMAE and MRMSE decreased by 19.50 and 12.08 respectively.Secondly,aiming at the problem of frequent domain migration in the underwater environment,this paper proposes a sea food identification and counting method for underwater image domain migration.Using the idea of generative confrontation network,apply it to the field of underwater image processing,and then construct the UGAN model,and then use the paired underwater images to train the UGAN network,let it generate high-quality images,and combine the existing The sea foods identification and counting method uses UGAN as an image preprocessing method,so that the model and the image to be detected are in the same environment,thereby improving the counting accuracy.The experimental results show that the improved average counting accuracy of the method proposed in this paper is 66.43% higher than that of the original model method.Finally,according to the needs of actual production work,on the basis of the identification and counting of sea foods that have been completed above,a sea foods identification and counting system is realized.This paper first designs the architecture of the sea foods identification and counting system,then builds relevant functional modules according to actual production needs,and then demonstrates the effect of the sea foods identification and counting system through specific cases,and verifies the effectiveness and usability of the sea foods identification and counting system,and finally complete the task of identifying and counting sea foods.Experimental results show that the system can meet the needs of actual production,and can complete the user’s demand for sea treasure inventory estimation.