Research on Multimodal Feature Selection of Rice Seed Quality Inspection

Abstract

Rice is one of the most essential foods in China. The quality of rice seeds affects the efficiency of rice production. High-quality rice seeds increase the yield and quality of rice and improve its nutritional value and taste. This study proposes a method for automatic inspection of rice seed quality based on RGB images and point clouds, aiming to provide better inspection of high-quality rice seeds based on a multimodal pre-fusion dataset. Eight widely distributed rice varieties in China were selected. A light field camera collected RGB images and point clouds of rice seeds. Two-dimensional and three-dimensional features were extracted, and a raw multimodal pre-fusion feature dataset was constructed. This study proposed a Comprehensive Evaluation Method to select features based on three feature selection methods (Chi-square Test, Minimum Redundancy Maximum Relevance, and Analysis of Variance). Three classifiers, Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Decision Tree (DT), were used to inspect rice seed quality. The accuracy based on the raw multimodal pre-fusion dataset was up to 96.8%. With four feature selection methods, the models' accuracy improved by an average of 3.1%, resulting in a final accuracy of 97.9%. The Comprehensive Evaluation Method for four rice varieties improves accuracy at 2.4%, 3.2%, 0.6%, and 5.7%, respectively. Moreover, the number of features required for best accuracy dropped to around 10 under feature selection. This study concludes that feature selection for multimodal pre-fusion dataset can improve the accuracy of rice seed quality inspection.

Keywords: rice seed, Quality inspection, multimodal pre-fusion, Feature Selection, machine vision