Research On Fish Species Recognition Based On Deep Convolutional Neural Network

China is one of the world’s largest aquatic countries,and fish resources are extremely valuable as aquatic biological resources.Aquatic biological survey projects often require a large number of accurate and efficient fish species identification work.Most of the existing fish identification methods use artificial naked eyes to identify features,and then manually check the search table for identification.This work requires experts who are familiar with fish classification and experience.The traditional fish species identification is mainly through identifying the unique morphological characteristics of the species,comparing the search tables,and identifying the species.This method of comparing search tables has certain requirements for taxonomy and identification experience,and requires extensive knowledge of fish taxonomy.In addition,the efficiency of manual identification is limited,and accuracy is easily disturbed by factors such as sensory fatigue,surrounding environment,and physical condition.Therefore,a low-cost intelligent tool for fish species identification is urgently needed.At present,the existing methods have certain defects in the generalization ability of different individual photos of the same type.Other methods need to mark the target of the fish main body,and cannot automatically learn the characteristics of the fish main body.The general applicability needs to be improved.Other methods cannot automatically adjust the learning rate and require manual tuning,which may cause the training results to fail to converge.In response to the above problems,this study takes 9common fish species in Chongming Island of Shanghai as the research object,and uses the Res Net50 pre-training model to train and recognize fish images respectively.First,collect and produce 9 kinds of fish images,including: Chinese carp,crucian carp,grass carp,wheat ear fish,bream,sand pond snake,Ziling goby,red eye trout,yellow catfish,etc.,a total of 7889 cameras data set.First use imgaug to augment the data,which is also a priori measure to reduce the fitting;then load Res Net50 pre-trained model parameters to initialize the neural network weight parameters training,reduce training time,and respond to the gradient by skipping the structure Disappear,deepen the network depth,improve accuracy and add global average pool and other methods for training and optimization.It solves the problem of too large models and too many parameters.After the training,a neural network classifier with good effect is obtained.After the data is rich,train to increase the classification category.Based on previous studies and considering practical fish scenarios,this paper selectively uses Res Net50 convolutional neural network with high recognition accuracy,small model size,and deep network depth as a pre-trained model network for fine-tuning.Convolutional network model with strong generalization ability,based on the characteristics of deep learning itself,free of image preprocessing,manually labeling the main target of fish;this paper uses data augmentation to enhance the original data set,and overcomes the early stage of some deep learning models The difficulty of "feeding" a large amount of data to the model;in the case of limited samples and limited computing resources,this paper adds random neuron discarding and global maximum pooling layer methods,using multiple layers and multiple small-size convolutional layers 1.Batch specification layer of standardized mapping,novel jump connection structure,reduce the number of channels of convolution layer,reduce the amount of parameters,improve calculation efficiency,and reduce overfitting;Adam adaptive learning rate based on second-order momentum Adjustment,to exclude obstacles such as manual model adjustment and learning rate may appear in the training process model oscillation,training results can not converge.At the same time,compared with the support vector machine based on the predecessors,based on the surrounding sound source scattering data as a factor,and the fusion factor fusion fish classifier,the classification accuracy rate is 92%,the training process is cumbersome and complex,and the accuracy rate is more Have better performance.Compared with the previous Inception V3 network model and SVM classifier,the model in this paper has a large advantage in model size and parameter calculation.The experimental results show that the fine-tuning convolutional network model based on the Res Net residual network in this paper has a higher network depth than the currently popular CNN network model,and the accuracy of the network increases synchronously.A Dropout layer and a maximum global pooling layer are added.Reduce the calculation of model parameters,use adam’s optimizer to greatly improve the convergence speed,and the model iteration rate is strengthened.The average accuracy of the network model in this paper is 96.56%,bringing higher calculation efficiency and accuracy than VGG network and Goog Le Net.Degree,achieved a good classification effect.Classification results: Res Net50 model achieved 96.5% recognition accuracy for each fish after training.The deep learning recognition system completed by this method has great advantages in model size and network depth.It is used in the field of fish image recognition in aquatic biology survey projects,and the research of fish image recognition technology based on deep learning is not only It can overcome the problems of traditional methods in feature extraction,and also improve the recognition accuracy,expand the recognition range,and play an active role in the recognition of aquatic animals in China.