Automated Tobacco Leaf Grading System Based on Deep Learning

To ensure the quality of tobacco products, tobacco leaves need to be graded to prevent mixing different quality tobacco leaves. The considerable intra-class variation and slight inter-class variation of different tobacco leaves may cause grading problems. Traditional methods of tobacco leaf grading rely on manual labor, which is inefficient and subjective. This work proposes an automated tobacco grading system based on deep learning. The reflectance image reflects information such as the texture, color, and shape of the tobacco leaf, and the transmittance image reflects the thickness of the tobacco leaf. Therefore, we capture both reflection and transmittance images of tobacco leaves and apply convolutional neural networks (CNNs) for automatic grading. During the acquisition process, the transmittance images of different areas are quickly stitched together to build a complete transmittance image of the tobacco leaf. Then the reflectance and transmittance images are aligned to improve the feature extraction of the CNN. After comparing the state-of-the-art models, we used DenseNet121 to grade the tobacco leaves and analyzed the effect of transmittance images on tobacco leaf grading.