A Deep Convolutional Neural Network Approach to Rice Grain Purity Analysis

Traditional rice grain classification is costly, time-consuming and requires sophisticated human expertise. Besides, computer vision based methods are still based on predefined morphological features that are often not transferable across different types of grains. In this paper, the feasibility of automated feature extraction for rice grain purity analysis has been demonstrated using a Convolutional Neural Network (CNN) based deep learning approach. Due to the lack of benchmark datasets, the paper defines a dataset with technician-verified, labeled images of different types of rice grains with a background of uniform illumination. Moreover, the paper also proposes the architecture of a CNN for automated rice grain feature extraction. The performance of a classifier trained on these features is compared to classifiers trained on morphological features used by modern computer vision approaches. It is found that in this dataset, the proposed method can detect the presence of native and foreign grains in a given sample of rice grains with superior accuracy which is at least 25% better in case of a multiclass classification scenario.