Nondestructive Detection Of Fruit Sugar Content Based On Spectral Reflectance Reconstruction Technology

As an essential ingredient in human diet,fruit is related to people’s dietary balance.In recent years,with the continuous improvement of living standards,consumers’requirements for fruit are no longer limited to"being able to eat",but rather requiring it to be"delicious",which puts forward higher requirements for the internal quality of fruit.Therefore,proposing a fast,non-destructive,and efficient non-destructive testing method for fruit sugar content has practical application value for today’s fruit industry and consumers.In this study,the spectral reflectance reconstruction technology based on RGB images was applied to non-destructive detection of fruit sugar content.An Android application software was developed to achieve data calculation and display,model storage and update.The effectiveness of this method for sugar content detection was verified by taking"Crown"pears and"Red Fuji"apples as examples.The main content of this study is as follows:(1)A spectral reflectance reconstruction algorithm based on fusion attention residual network was proposed.A deep residual network is built based on Tensor Flow.Average pooling operation is used to reduce the number of model parameters.Adam optimizer is used to replace the random gradient descent optimization algorithm.Root Mean Square Error(RMSE)is used as the loss function.To selectively highlight information region features and enhance spatial feature representation,a spatial attention mechanism has been introduced;To model the interdependence between intermediate feature mapping channels to adaptively calibrate channel feature responses and enhance feature learning,an improved channel attention module was introduced,and dynamic convolution was used to enhance the model’s expression ability.By comparing with algorithms proposed by other scholars on public datasets,the reconstruction spectrum based on fusion improved channel attention residual network proposed in this study achieved better accuracy on CAVE and NUS datasets,with RMSE of 2.4638 and 2.6230,respectively.Compared with the previous best reconstruction algorithm,RMSE was reduced by 1.4%and 1.9%,respectively;Compared to deep residual networks without fused attention mechanisms,the RMSE of reconstructed results on CAVE,ICVL,and NUS datasets decreased by 13.9%,17.1%,and 9.7%,respectively,and was superior to the reconstruction accuracy of residual networks with fused ordinary attention mechanisms.The results demonstrate that the algorithm proposed in this study can meet the needs of spectral reconstruction and has high accuracy.(2)Designed intelligent mobile application software to undertake functions such as data calculation and result display,and combined with My SQL relational database and Alibaba Cloud server technology to achieve cloud call and storage of data and models.Obtain sample RGB images by rewriting the CAMERA method and calling the camera,album,and other functions of the phone;To improve computing speed,the deep learning model is transplanted into the application,extracting RGB responses and completing spectral reflectance reconstruction using the data computing capabilities of smart phones;In order to display the reconstruction results in real time,a Chat View was designed to draw the reconstructed spectral curve.Complete the prediction of the target index values based on the reconstructed spectral reflectance and the sugar content prediction model stored in the cloud database;All forecast results can be saved to a local or cloud database.To facilitate software management,set administrator permissions in User Activity to modify local and cloud data.Non administrators only have read permissions.(3)Verified the effectiveness of spectral reflectance reconstruction technology based on RGB images for rapid and non-destructive detection of fruit sugar content.Taking two common fruits,"Crown"pear and"Red Fuji"apple,as examples,a fruit dataset was established,and compensation algorithms were designed for interference factors such as ambient light and fruit skin texture;To address the issue of pixel mismatch between smartphones and hyperspectral cameras,pixel registration is performed on RGB and hyperspectral images using the Forstner operator;Divide the training sample set based on the spectral physicochemical value co-occurrence distance algorithm;A sample holder suitable for smartphones was designed for liquid samples such as fruit juice.Different spectral reconstruction methods and sugar modeling methods were compared,and the results showed that the fusion attention residual network and partial least squares regression model can achieve the best prediction accuracy.For the"Crown"pear sample,the extreme error between the reconstructed spectral reflectance and the true spectral reflectance is 0.067.The R_p and RMSEP(Root Mean Square Error of Prediction)of the test set in the fruit sugar prediction results are 0.855 and 0.651°Brix,respectively.88.5%of the samples have prediction errors less than±1°Brix;In the prediction results of juice sugar content,R_p and RMSEP were 0.878 and 0.406°Brix,respectively,with 96.7%of the sample prediction errors less than±1°Brix.For the"Red Fuji"apple sample,the extreme error between the reconstructed spectral reflectance and the true spectral reflectance is 0.054.The R_p and RMSEP in the fruit and juice prediction results are 0.867,0.638°Brix,and 0.888,0.621°Brix,respectively.Among them,90.7%and 93.3%of the sample prediction errors are less than±1°Brix,respectively;The single detection time based on mobile application software is approximately 4～5 seconds.The above results indicate that the spectral reflectance reconstruction technology based on RGB images proposed in this study can be used for real-time detection of fruit sugar content.