Non-destructive Detection of Moldy Camellia Oleifera Seeds Based on Deep Learning and Hyperspectral Imaging

Camellia oleifera, a woody oil plant of high economic value, faces potential mold growth during storage and transportation, which could significantly impact its sales and processing. Therefore, it is crucial to identify and select moldy camellia oleifera seeds. Due to the fact that the mold growth of camellia oleifera seeds typically originates from the inside of the seeds, the mold is not visibly apparent on the surface of the shells in the early stages of mold growth. The existing chemical methods are time-consuming, labor-intensive, and result in sample damage. To address this challenge, this study introduces a non-destructive detection method for assessing the extent of mold growth in camellia oleifera seeds efficiently and effectively. The method is based on deep learning and hyperspectral imaging (HSI) technology. A hyperspectral imaging system was utilized to capture hyperspectral images of camellia oleifera seeds at three degrees of mold. In response to the limitations of traditional machine learning methods in hyperspectral classification, which only focus on learning spectral features and yield unsatisfactory results, this study introduces a deep learning approach based on three-dimensional convolution. This algorithm is capable of extracting features from both spatial and spectral information. The classification performance of the proposed model is compared with machine learning methods that rely solely on spectral information and deep learning methods that use only spatial information from visual images. Based on the dataset constructed in this study, experimental results demonstrate that the proposed method achieves an accuracy of 88.99%, surpassing other methods by at least 18.6%.