| China has always been one of the world’s largest fruit producers and consumers.The fruit industry is an important part of our national economy.However,although China’s fruit industry is huge,it has been in a large but not strong situation for a long time.The domestic supply of high-quality fruits is inadequate,most of them depend on imports,and there is a large trade deficit.The key reason is the China’s relatively backward fruit quality detection and grading technology.At present,most researches on non-destructive testing of fruit quality are based on the hyperspectral information of the fruit.Recent research shows that due to the fusion of the physical and chemical characteristics of fruit tissue structure,the fruit quality detection model constructed by fruit optical parameters has higher detection accuracy,but the fruit optical parameter collection method is complicated,and the fruit tissue structure needs to be destroyed.This paper proposes a method for non-destructive detection of fruit internal quality based on deep feature fusion,which combines the characteristics of the hyperspectral and optical parameters of the fruit for joint decision-making.It solves the problems of low precision of single-spectrum detection and difficult to obtain optical parameters,and provides new ideas for non-destructive detection of fruit internal quality.The main research contents and results of this paper are as follows:(1)Data acquisition:245 hyperspectral images,optical parameters and quality parameter information of Red Fuji Apple were collected,and the collected optical parameters and quality parameters were classified and processed.(2)Internal quality detection of apple based on 3D-CNN:the method of combining 3D-CNN and multi-task learning to achieve simultaneous detection of multiple internal quality parameters of apples based on hyperspectral,the classification accuracy of the model reaches:sugar content:93.97%,hardness:92.29%,water content:93.36%,joint:89.84%.The characteristics of the 3D hyperspectral image are extracted through this model,and regression analysis is performed with the quality parameters to obtain the correlation coefficient of the regression model:sugar content:0.827,hardness:0.755,water content:0.862.(3)Apple optical parameter inversion based on transfer learning:the Monte Carlo method is used to simulate the apple light distribution map to solve the problem that the apple optical parameters are difficult to obtain.On this basis,a transfer learning method based on the adversarial learning and distribution matching is proposed.A two-stream CNN is used.One network extracts the characteristics of the simulated apple light distribution map and classifies the optical parameter information of the simulated data.Another network extracts the characteristics of the measured apple brightness distribution map,and the two networks share parameters.In network migration,a method based on adversarial learning is introduced,using a domain discriminator to distinguish features from measured and simulated apple brightness maps,and a gradient inversion layer is used to reduce the domain features differences between measured and simulated apple brightness maps.Afterwards,a method based on distribution matching is introduced,and the maximum mean difference is used to further align the characteristic distribution of the measured and simulated apple brightness map.The classification accuracy of the optical parameter information retrieved by the migration model reaches:the peel absorption coefficient:73.47%,the pulp absorption coefficient:94.89%,the peel scattering coefficient:80.02%,and the pulp scattering coefficient:90.56%.(4)Internal quality detection of apple based on deep feature fusion:the extracted quality parameter features based on 3D hyperspectral images and optical parameter features based on 2D brightness images are subjected to deep feature fusion to improve the final detection precision of internal quality parameters.The classification accuracy of Apple’s internal quality parameter detection model established by the improved deep feature fusion network reaches:sugar content:95.02%,hardness:93.47%,water content:96.29%,and joint:91.38%.Using this model to extract the characteristics of apple hyperspectral data,and regression analysis with the quality parameters,the correlation coefficient of the regression model obtains:sugar degree:0.887,hardness:0.814,water content:0.891.The results prove the superiority of the proposed internal quality detection method based on deep feature fusion. |
