A method for constructing optical detection model of wheat seed purity based on sample generation and contrastive learning strategy

Abstract

Seed purity is an essential indicator of seed quality, and use of optical techniques for seed purity testing has been extensively reported. Currently, most optical detection models for seed purity rely on the assumptions that the training set used to develop detection model contains all possible impurity varieties, and that the number of positive (the variety of interest) and negative samples (all possible impurities contained in that variety) in the training set is balanced. This assumption is often difficult to fulfil in practice due to the complexity of the seed production process. To address this issue, an optical detection model construction method for wheat seed purity based on sample generation and contrastive learning strategy was proposed in this study. The proposed method firstly employed a linear generation strategy to expand the impurity samples in training set, thus enhancing the diversity of impurity samples and improving the unbalance between the positive and negative samples. Thereafter, the contrastive learning loss function was introduced to train the optical detection model based on the deep convolutional neural network, so as to enhance the feature differences between positive and negative samples, and to improve the model's recognition accuracy for unknown impurity samples. Hyperspectral images of 4200 wheat seeds from six varieties were acquired, and a purity detection model for each variety was built by using the average spectra of the seed as input. The proposed method achieved average accuracy of 95.33 %, with improvements of 21.18 % over SVDD (78.67 %), 12.13 % over LSSVM (85.02 %), and 13.06 % over CNN (84.2 %), respectively. Further studies shown that the proposed method still maintains good detection accuracy under the condition that the test set contains multiply unknown impurities. The proposed method provides a feasible way for the construction of an optical detection model for seed purity in a real scenario.