The Research Of Mutton Adulteration Detection Based On Deep Learning Used Smart Phone

Mutton,with its low cholesterol and high protein content as well as its rich nutritional value and delightful taste,is a significant part of the Chinese meat market.In recent years,its price has been escalating steadily,far exceeding that of pork and other meats.Some illegal merchants have been adulterating pork into mutton in order to obtain greater economic benefits.At the same time,in order to make the adulterated mutton look and smell more like mutton,and to achieve the effect of"deceiving the public",illegal traders add mutton flavor essence and dyeing agents to the adulterated mutton.Mutton adulteration with pork under the effect of mutton flavor essence and dye not only seriously infringes on the economic interests of consumers and disrupts market order,but also poses a threat to the health of consumers,thus triggering food safety issues.Therefore,a rapid and accurate detection technique needs to be developed to detect adulterated mutton under the effect of mutton flavor essence and dye.This thesis takes the adulterated mutton under the effect of mutton flavor essence and dye,pork and mutton.As the research object,shooting the images of the samples by intelligent mobile phone,establishes the qualitative classfication model and quantitative prediction model of adulterated mutton under the effect of mutton flavor essence and dye respectively by deep learning method,compares the accuracy,R~2 and RMSE of the model and selects the best model.Then transplanting the selected deep learning model to the mobile to develop a mobile APP for detecting adulterated mutton under the effect of mutton flavor essence and dye.The main research contents include:(1)Research on the establishment of a qualitative identification CNN model using mobile image data of pork adulterated mutton under the effect of mutton flavor essence and dye samples,pure mutton samples and pure pork samples.Replacing the original residual structure in the Res Net network framework with an inverted residual structure to learn higher-level semantic features,making the model more lightweight,using attention mechanism to strengthen the difference of different meat features,and improving the model accuracy.The classification accuracy of CBAM-Invert-Res Net50 model for the datasets of loin,fore shank and hind shank under the effect of mutton flavor essence and dye were 95.19%,94.29%and 95.81%,respectively.For the more complex three-part mixed dataset,feature fusion combined with transfer learning method can be used to achieve the complementary advantages of multiple features,obtaining more robust and accurate recognition results.After feature fusion,the classification accuracy of CBAM-Invert-Res Net50on the three-part mixed dataset is 88.50%.(2)Research on the establishment of quantitative prediction CNN model based on mobile image data of adulterated mutton with different pork part and different proportion under the effect of mutton flavor essence and dye samples.The R~2 of the CBAM-Invert-Res Net50 model for predicting the dataset of adulterated mutton loin,fore shank and hind shank respectively was 0.9373,0.8876 and 0.9055,and the RMSE was0.0268,0.0357 and 0.0316 respectively.After feature fusion,the R~2 and RMSE of CBAM-Invert-Res Net50for three-parts mixed adulterated mutton dataset prediction are 0.0589 and 0.0220,respectively,Compared with before fusion,R~2 increased by 0.0325,and RMSE decreased by 0.0070.(3)Based on the classification model and prediction model optimized in(1)and(2),the design and performance test of mutton adulteration detection software under the action of additives based on smart phones were carried out.Tensorflow deep learning framework was used to deploy the model to the mobile terminal,and Android Studio development tools were used to design the front-end interface of the detection software and program the image preprocessing function,model call function and mutton adulteration detection function.Carry out functional testing,testing time testing and precision testing of software.The average detection time of mutton adulteration detection software was 0.5 s,the classification accuracy was88.33%,the content prediction R2 and RMSE were 0.8995 and 0.0327,respectively.