Fish Sign Recognition System Based On Machine Vision

In recent years,the production of aquatic products has been increasing day by day,and the development of marine fishery resources has received more and more attention from the country.The problem to be solved is how to deal with the fish after fishing in a more intelligent and automatic way.The classical fishing and classification of aquatic products is mostly an extensive assembly line model,which is mainly based on manual observation and supplemented by simple machine operation.With the increase of labor costs caused by the aging of the population,the traditional detection and classification methods are faced with the problems of large labor consumption and low efficiency,and the subsequent sales costs are also increasing gradually.In the wake of revolutions in science and technology and the emergence of computer-aided engineering,machine learning has gradually occupied an important position in the realm of image processing.It has taken centre stage in the main current pattern of object detection and classification.Based on this research background,there are two key issues to be researched and solved in this paper.The first problem to be solved urgently is the efficiency of object detection.The work of the system and the procedure of image recognition will spend a lot of time and space.It is significant to apply a functional algorithm model for the sake of achieving the real-time record of the system.The second problem is the accuracy of object detection.The traditional fish sorting system often selects the conveyor belt for resource transmission,but the types of targets to be detected are diverse,with different postures and fast moving speed.The algorithm needs high reliability and stability to process the collected pictures,and control the error rate within a small range to guarantee the correctness of the output parameters.In the context of machine vision background,this system structures a fish sign recognition system based on machine vision to substitute manual labor,so as to capture real-time images,collect fish sign data and classify them with algorithm.At the beginning,this paper introduces the research status and application scenarios,sums up the burgeoning demand of fishery development and image processing technology,introduces the research progress of machine vision in object detection all over the world,and then designs the system framework according to the actual industrial environment.For the sake of addressing the above two problems,two different methods are proposed separately in this paper.The optimized algorithm model is analyzed in detail accordingly.The sensor network was established,the simulation experiment platform was built,and the data demonstration and analysis under different models were carried out.The primary work contents are as follows:(1)To solve the speed problem of fish target detection,aiming at the difficult point that the classical Canny algorithm for edge detection is easy to be influenced by noise and the edge is easy to be misdetected,nonlinear filtering is used to preserve edges effectively to take place of the Gaussian filter and the Sobel operator gradient template is enhanced.The improved algorithm is used to detect the edge of the object and extract the contour,so as to obtain the real size of fish.(2)To solve the fish image recognition accuracy problem,Faster R-CNN deep learning algorithm is used and VGG16 is used as a feature extraction network to enhance the accuracy of target recognition.However,VGG16 has large network model parameters and slow model training speed.This paper aims at using Mobilenet V2 as backbone feature extraction network to substitute VGG16,which will slightly impact the accuracy,decrease the training parameters of the model and raise the speed of system operation.(3)The hardware platform was built to simulate the real industrial sorting environment.The classical Canny algorithm and the optimized Canny algorithm model,VGG16 as the feature extraction network and Mobilenet V2 as the feature extraction network,and the measurement results before and after data enhancement were demonstrated and analyzed in detail.The advantages and disadvantages of the algorithm were analyzed from different dimensions.Finally,the recognition error rate of the improved Canny algorithm is 4.62%,which is 3.38% higher than the classical measurement result.The average measurement error rate of VGG16 network was2.59%,and the average measurement error rate of the Mobilenet V2 network was3.19%.Mobilenet V2 reduced the network training parameters with only 0.6% accuracy loss,and its efficiency and accuracy met the requirements of the recognition system.