Research on Small Sample Apple Defect Classification Method Based on Siamese Network

To address the issue of collecting a vast amount of labeled samples in traditional deep learning methods for apple defect classification, a metric learning-based method (Siamese Network model) is proposed in this paper for apple defect classification with few samples for training. Two ResNet-18 sub-networks with shared weights are used as feature extractors, and two apple images are formed into a sample pair at a time and input into the sub-network to obtain their respective high-dimensional feature vector representations, the Euclidean distance between them is calculated and the classes they belong to is determined based on the similarity. The dataset used for the experiments contains 550 labeled apple images (two categories: defective and non-defective), 330 images are allocated for training, 110 for validation, and 110 for testing. The results show that Siamese Network achieves better results than many classical models with few-sample datasets, with an accuracy of 91.82% and an AUC value of 0.984 on the test set, which validates the effectiveness of the method in this paper for apple defect classification with few samples.