Fresh grape quality inspection based on dual-channel improved convolutional neural network

ObjectiveTo achieve accurate and non-destructive testing of key quality indicators of fresh grapes, we proposed a dual-channel improved convolutional neural network (CNN) method for fresh grape quality inspection.MethodsA fiber-optic spectrometer and a CCD camera were used to collect visible near-infrared spectra and characterize the image information of grape samples. The dual channel improved CNN model was established and the Gramian angular field (GAF) was used to convert one-dimensional spectral data into two-dimensional images for the extraction of effective features from spectra data. The convolutional layers with different sizes of convolution kernels were designed to extract features from the converted spectral two-dimensional images and grape characterization images, thus enhancing the comprehensive perception ability of the CNN model. On this basis, the dropout method was used for the fully connected layer to reduce and fuse the extracted features, ultimately achieving accurate prediction of key indicators of grape quality.ResultsCompared with other three grape quality inspecting methods, the CNN method showed the root mean square error decreases of 50.48%, 57.44%, and 49.56% and the correlation coefficient increases of 4.89%, 3.13%, and 2.17%, respectively.ConclusionThe dual-channel improved CNN method can achieve non-destructive testing of key indicators of grape quality