Discrimination of Tomato Maturity Using Hyperspectral Imaging Combined with Graph-Based Semi-supervised Method Considering Class Probability Information

Accurate and nondestructive maturity discrimination of tomatoes is significant for harvest and preservation. Color differences between intermediate adjacent maturity stages are not significant, so it is time-consuming to obtain a large amount of precise maturity labels and difficult to discriminate multiple maturity stages by visual features merely. This study designed a semi-supervised method based on hyperspectral imaging technology for tomato maturity discrimination by using a small amount of labeled samples. Firstly, the hyperspectral data of tomato samples was extracted by the hyperspectral imaging system and pre-processed by the multiple scattering correction algorithm. Then, the class probability information of unlabeled samples was described by the sparse coding of labeled samples. Next, a semi-supervised algorithm based on Laplacian score and spectral information divergence (SIDLS), which used class probability information to construct graphs, was designed to select a representative waveband subset. Finally, the sparse representation model based on class probability information (CSR) was established to construct a connection graph, and label propagation algorithm was used to discriminate tomato maturity. Experimental results demonstrated that SIDLS algorithm had an advantage in the feature selection over semi-supervised algorithm based on Laplacian score (SSLS) and Semi_Fisher Score (SFS) algorithm, and CSR model was also superior to other graph construction methods in constructing a more discriminative graph. The discrimination accuracy of this method could reach 96.78%, which shows a promising prospect in tomato maturity discrimination.

Automated summary

A semi-supervised method based on hyperspectral imaging technology was designed for tomato maturity discrimination using a small amount of labeled samples, showing promising prospects with a discrimination accuracy of 96.78%.