Abstract: A nondestructive examination was conducted to identify unsound kernels by using hyperspectral imaging technology. Based on 932 wheat samples including 486 normal samples, 170 damaged samples, 149 worm-eaten samples and 127 black germ kernel samples, a hyperspectral image acquisition system was used for collecting hyperspectral information, and then a convolutional neural network (CNN) was established based on 30 wavebands selected from 116 wavebands for each sample. The CNN model comprised two convolution layers. The first layer consisted of 32 convolution kernels (3 × 3) and the second layer consisted of 64 convolution kernels (5 × 5). The pooling layer was developed with the maximum pool. The activation function was developed with rectified linear units (ReLu). To avoid overfitting, a dropout layer was linked to the fully connected layer, and the parameter was set as 0.5. When other parameters remained default, recognition rates for the calibration and test sets were 100.00% and 99.98% respectively. Finally, a support vector machine (SVM) model was built and compared with the CNN model. The SVM model developed with 90 wavebands selected from 116 wavebands showed a recognition rate of 94.73% for the test set. The recognition rate of the CNN model was better than that of the SVM model. Thus, this research showed that the CNN model allowed for accurate, rapid and nondestructive detection of unsound wheat kernels.

Key words: wheat, unsound kernels, hyperspectral detection, CNN model