Research On Fish Species Recognition Algorithm Based On Machine Vision

Marine and inland fisheries,including aquaculture,are a source of income for some 820 million people around the world which involve harvesting,processing,marketing and distribution activities.As a major power in aquaculture,China is in urgent need of solving the key problems in the future development of fishery industry by means of modern information technology.The identification of fish species can provide favorable technical support for the fishery industry,which has important application value for aquatic product breeding,fishery processing industry,fish resource investigation and ecological protection.The identification of fish species is mainly faced with two problems:firstly,the traditional use of manual identification will cause the problems of high labor intensity and low recognition rate,and there are few experts in fishery identification,so it takes a long time to train novices,or even longer time to establish practical experience;Secondly,fish recognition is also affected by various factors such as light intensity,background habitat and different species.In order to improve the recognition rate of fish species under complex background,this paper deeply studies the method of fish species recognition.The main research contents and results are as follows:1.Aiming at the problem of fish species recognition under complex background,this paper proposes a multi-feature fish classification algorithm based on particle swarm optimization SVM.In this method,fish images in natural environment were used for experiments.First extracted color,gradient direction histogram(HOG)and gray level co-occurrence matrix(GLCM)feature form feature vector from the original image,and put forward the method to choose the optimal weight ratio of feature fusion,then used the PCA technology to extract feature vector for dimension reduction to eliminate redundant data,finally proposed that use the particle swarm optimization to identify the classification of the support vector machine(SVM)classification method.The experiment proves that the application of HOG feature to the field of fish classification can significantly improve the classification effect compared with the traditional method of extracting fish features.After using particle swarm optimization support vector machine(SVM)classification of fish,use natural environment under the complex background of fish experiment database,compared to traditional support vector machine(SVM)classification method can accurately classify fish,the highest classification accuracy of 94.7%,the method can be applied to the actual fish image data sets,implement effective monitoring of fish biodiversity.2.The above method has some advantages for the classification of a small number of fish species,but it has some limitations for the comparison of more fish species.In view of the situation of numerous Marine fish species and high similarity.This paper proposed a fish species recognition method based on deep learning.Deep learning methods could effectively identify fish at different resolutions even if there was a strong textured background in the image.The network structure designed Block module to extract features.In the process of convolution,the Block module adopted the idea of void convolution to expand the receptive field and capture multi-scale context information.Adding the attention mechanism and improving the performance of the neural network could make the model pay more attention to the target object and extract the fish features more effectively.The network design carried out extensive network optimization to determine the optimal model.The model experimented on WildFish and Fish4Knowlege data set and compared experiment with the deep neural network model related to image recognition.These classification recognition models included:VGG,ALexNet,SqueezeNet DenseNet and ResNet models.Experimental results showed that the proposed network model had certain advantages in classification accuracy.The highest classification accuracy was 97.5 percent.