| How to realize the semantic representation of fish feeding desire throughout the feeding process is not only a challenge for production management in the recirculating aquaculture system(RAS),but also a key scientific problem for achieving precision feeding and welfare farming.This paper explored key technologies for a variety of challenges and pain points in the feeding field of RAS from the production practice,using Micropterus salmoides as the research object.A semantic representation model of Micropterus salmoides feeding desire based on photoacoustic coupling was created using computer vision,audio information collecting and processing,and deep learning technologies.The principal research findings and contents are as follows:(1)A method for assessing the fish feeding intensity of fish based on spatiotemporal behavior characteristics was proposed.This method combined the improved kinetic energy model and customized recurrent neural network,and was applied to the Micropterus salmoides feeding intensity assessment task in RAS.The average classification accuracy(i.e.,the assessment accuracy)reaches 98.31%.In order to verify the effectiveness of the model,the evaluation accuracy of this method was compared with that of the fish appetite grading method based on CNN model(Le Net5)and the fish appetite grading method based on OFA-LSTM(combination of optical flow algorithm and long short term memory network),and the accuracy of the proposed method was improved by 23.44% and 3.5% respectively,compared with the above two methods.In addition,to evaluate the feasibility of the model,the performance of the proposed fish feeding intensity assessment method based on spatiotemporal behavior characteristics was evaluated under different datasets,and it was determined that the accuracy rate on the early feeding stage dataset was 97.08%,while the accuracy rate on the mid feeding stage dataset was 97.35%,which was only 1.23% and 0.96% lower than the accuracy rate on the whole feeding stage dataset.It was confirmed that the proposed method can accurately evaluate the fish appetite under different datasets and the feasibility of the method was also verified.(2)In light of the time dependence of the existing fish feeding desire grading methods on time data,a fish feeding desire representation method based on limited data and efficient use of spatial features of fish feeding behavior was proposed in order to realize the efficient use of spatial features of fish feeding behavior.Due to the construction of the improved kinetic energy characteristic map and the customization of the graph convolution network structure,the research results indicate that our method has a greater capacity for learning than the conventional method for assessing fish appetite,particularly for time-limited feeding behavior data.Using only the first4.2 seconds of the input data(accuracy: 98.66%,recall rate: 98.59%,F1 score: 98.62%)and the first 8.3 seconds(accuracy: 98.61%,recall rate: 98.63%,F1 score: 98.62%),the accuracy rate of appetite grading can reach 98.60%,which is little different from the accuracy rate of 98.89% of the input data of 25 seconds(accuracy: 98.92%,recall rate: 98.90%,F1 score: 98.90%).(3)In view of the fact that the method for extracting feeding behavior features based on computer vision is susceptible to the interference from light and other factors,a method for recognizing and grading the feeding intensity of fish based on audio features was proposed.This method extracted the acoustic information during the feeding process of fish and converts it into time-frequency map using Fast Fourier Transform algorithm(FFT).Finally,the improved depth residual shrinkage network(DRSN-16)was established to realize the recognition and classification of fish feeding intensity.The results showed that the accuracy rate of the proposed DRSN-16 method for grading the fish feeding intensity was 99.28%,and the average accuracy rate,average recall rate and F1 score were 99.34%,99.48% and 99.41%,respectively.Compared with the traditional CNN models(Res Net18,Res Net34,VGG11,VGG13,VGG16,VGG19,Mobile Net,Mobile Net v2,Shuffle Net and Shuffle Net v2),its effect is improved to some extent,and it also has some advantages in terms of computing complexity(model parameter quantity).In addition,by comparing the accuracy of the proposed method under ELU,CELU,LRe LU,and Re LU activation functions,it is concluded that the model has the best grading effect on fish feeding intensity under Re LU activation functions.The aforementioned results have demonstrated the feasibility of the application of audio information of fish feeding behavior in RAS.(4)Aiming at the deficiency of using only the optical or acoustic characteristics of fish feeding behavior to represent the feeding desire in RAS,a semantic representation model of fish feeding desire based on photoacoustic coupling characteristics was proposed.From the perspective of actual production,the model fully exploits and utilizes the optical and acoustic features of fish feeding behavior by using deep learning multimodal fusion technology,and realizes the complementary advantages of photoacoustic features.The research findings indicate that the combination of FCN and LSTM can completely learn the optical characteristics of fish feeding behavior,while Resnet has great advantages in the extraction of acoustic features.In this study,FCN-LSTM and Res Net methods were coupled to fuse and utilize the photoacoustic features of fish feeding behavior.The accuracy of feeding desire representation is 99.81% by multiplicative fusion.The performance of the model was evaluated using several optimizers.In addition,the dataset was constructed using the feeding test data of three specifications of Micropterus salmoides with body length of 7～8 cm,13～14 cm,and 19～20 cm,respectively,in order to test the grading effect of the model on the feeding desire of Micropterus salmoides of varying body length.The results demonstrate the effectiveness of the model and reflect the feasibility of the model applied in different Micropterus salmoides culture densities. |
PDF Full Text Request