Research On The Recognition Of Shellfish Based On Convolutional Neural Network

Each shellfish has its own unique shell shape and color characteristics,but the overall appearance is similar.For ordinary people,even professionals,it is difficult to accurately identify the category of shellfish through visual observation.If the recognition is carried out through experiments,it will take too long to realize real-time recognition.Therefore,this paper studies the shellfish recognition method based on Convolutional Neural Network（CNN）,and develops a shellfish recognition mobile APP,so that users can easily identify the shellfish around them online at any time,and organically combine"recognize shellfish,appreciate shellfish,and love shellfish"to arouse the public’s love,and stimulate the public’s awareness of protecting marine life.The research work of this paper mainly includes the following four aspects:（1）The image dataset of shellfish is established.We successively collected shellfish from 9 coastal cities in China and constructed 2 datasets:Single-target Shellfish Recognition（SSR）dataset,Multi-target Shellfish Recognition（MSR）dataset.Two datasets are suitable for single-target shellfish recognition and multi-target shellfish detection and recognition,respectively.The SSR dataset contains 93,574images of 68 categories,and the MSR dataset contains 70 categories of 7,573 images.The methods proposed in this paper have been trained and verified on these two original datasets.The datasets are collected and established under the guidance of professionals,which ensures the accuracy and reliability of the datasets in this paper.（2）A single-target shellfish recognition method based on CNN is proposed.Difficulties in single-target shellfish recognition:Firstly,it is difficult to identify shellfish with high similarity;secondly,the imbalance of different shellfish samples increases the difficulty of recognition.In order to solve the above problems,a single-target shellfish recognition method based on CNN is proposed,which aims to improve the expression ability of CNN for shellfish and improve the recognition accuracy.Firstly,for the classification problem of shellfish with similar characteristics,a filter information metric including output entropy metric and orthogonality metric is proposed;secondly,driven by the filter information metric,a filter pruning and repairing model is proposed to improve the feature expression ability of CNN for useful information;finally,on the basis of filter pruning and repairing,for the problem of unbalanced shellfish classification,a hybrid loss function including regularization term and focal loss term is proposed to improve the objective function of shellfish classification,and the accuracy of shellfish classification is improved.The method was experimentally verified on the established SSR dataset,and the accuracy rate of shellfish classification is 93.95%,which improved the classification accuracy rate of the benchmark network（VGG16）by 13.68%,compared with the network Alex Net,Google Net,Res Net50,SN＿Net,Mutual Net and Res Ne St improve the accuracy of shellfish classification by 0.46%,17.41%,17.36%,4.46%,1.67%and 1.03%respectively.The experimental results prove the effectiveness of the method.（3）A multi-target shellfish detection and recognition method based on target detection network is proposed.Difficulties in multi-target shellfish recognition:Firstly,it is easy to confuse the same shellfish detection;secondly,when there are shellfish with large differences in size in a picture,the recognition accuracy of smaller shellfish is low.In order to solve the above problems,a multi-target shellfish detection and recognition method based on target detection network is proposed to improve the recognition accuracy.Firstly,an improved Att Res Net101 backbone network model based on Res Net101 is proposed;secondly,an improved feature extraction model is proposed;finally,based on the improvement of the backbone network and the feature extraction model,an improved multi-target shellfish detection method based on the target detection network is proposed.The accurate detection of multi-target shellfish is realized,and the detection accuracy of small shellfish targets is improved.The method was experimentally verified on the established MSR dataset,and the value of AP₅₀reached 94.585%,the value of AR₁₀ reached 87.3%.Compared with FCOS,OTA,DeFCN（fcos+poto）and DeFCN（fcos+poto+3dmf）,the method has better detection efficiency for shellfish.The accuracy rates of AP₅₀,AP₇₅,m AP,AP_S,AP_M and AP_Lhave all improved,of which the values of AP₅₀ have increased by 2.374%,1.865%,0.992%and 0.644%respectively;the values of AP_S have increased by 7.476%,15.149%and 19.901%and 10%respectively.The experimental results prove the effectiveness of the method.（4）A shellfish recognition system based on deep learning is developed.Based on the single-target shellfish recognition method and the multi-target shellfish detection and recognition method,the intelligent recognition of shellfish targets is realized by establishing the interaction between the mobile client and the server,and the corresponding shellfish popular science information is displayed in the shellfish recognition mobile APP according to the recognition results.The system is tested and is already in the trial operation stage.