Nondestructive classification of internal defects in camellia seeds and chestnuts using X-ray imaging: optimization and comparison of deep learning classifiers

Camellia seeds and chestnuts are two important nut resources that provide healthy and nutritious oils and foods for humans. However, frequently occurring internal defects such as insect-damaged, mold-infested, and withered cause food safety risks and loss of economic value. In this study, X-ray digital radiography (DR) imaging was used in conjunction with deep learning (VGG16, ResNet18 and DenseNet121) to classify internal defects. The number of pre-processed images was increased to 10 times, reaching 2040 for camellia seeds and 2370 for chestnuts, through the use of image enhancement methods. Subsequently, these images were divided into training, validation, and prediction sets in a ratio of approximately 70:15:15 to construct and evaluate deep learning classifiers. A hyperparameter optimization strategy was developed to search for optimal hyperparameter combinations for training classifiers.

In addition, the performance and consistency of the three deep learning classifiers were compared in terms of sensitivity, specificity and accuracy using the outputs of the prediction sets. The best classifier DenseNet121 obtained a satisfactory area under the curve (AUC) of 0.9991 for classification of camellia seeds, while ResNet18 gained the best result for chestnuts with an AUC of 0.9234. This study integrates low-energy X-ray DR with deep learning to non-destructively classify internal defects in camellia seeds and chestnuts. It circumvents limitations of molecular spectroscopy and CT imaging (e.g., penetration depth, cost, reconstruction time) while offering actionable guidelines for industrial-scale implementation in food quality control.