Rice Seed Images Recognition Based on HOG Descriptor and Trimming Imputation Approach

Abstract

Rice seed classification is one of the most popular research topics in countries with the agricultural economy. The intention of the classification stage is to ensure the purity of rice seeds. In this paper, we propose Histogram of Oriented Gradients (HOG) descriptor to tackle this task. HOG is one of the most robust features that has a wide range of application in computer vision task. However, in practice, a problem appears after we extracted HOG descriptors from dataset contains rice seed images with different sizes. The problem is the extracted feature vectors have different numbers of dimension. This limit of HOG descriptor causes the classification stage to become impossible. As the current solutions are not optimal enough, we propose combine the HOG descriptor with missing values imputation method in order to help the obtained feature vectors gain the same number of dimensions. The proposed approach is then applied on VNRICE benchmark dataset and evaluated by KNN classifier.