Detection of impurities in wheat using terahertz spectral imaging and convolutional neural networks

The aim of this work was to propose a method to rapidly and effectively detect impurities contained in wheat based on a combination of terahertz spectral imaging and a convolutional neural network. First, the spectral characteristics of wheat, wheat husk, wheat straw, wheat leaf, wheat grain, weed, and ladybugs within the range of 0.2–1.6 THz were studied using the THz spectral imaging, and the corresponding frequency-domain spectra were obtained using Fourier transformation. The absorption coefficient and refractive index were then calculated. THz pseudo-color imaging was conducted next on wheat and its impurities according to the principle of maximum frequency-domain imaging, and a novel Wheat-V2 convolutional neutral network (CNN) was designed to extract the data and information regarding spectral imaging features. Finally, the designed Wheat-V2 model was compared with the ResNet-V2_50 and ResNet-V2_101 models under the same conditions. In addition, the loss function and confusion matrix indicators were used to evaluate the experimental results. The results show that the designed Wheat-V2 model can effectively recognize the impurities in wheat images, with a recognition accuracy of 97.56% and 98.58% for the verification sets Top_1 and Top_5, respectively. In addition, the designed Wheat-V2 model achieved an average F1-score of 97.83% in terms of image recognition of various impurities, which is higher than that achieved by conventional models, i.e. ResNet-V2_50 and ResNet-V2_101. This indicates that the method combining THz spectral imaging and CNN can be used for the detection of impurities in wheat. In addition, the results also indicate the potential of application of CNN in THz imaging detection of impurities in wheat, providing a nondestructive testing method for the recognition of impurities in other grains.