Machine learnt image processing to predict weight and size of rice kernels

Abstract

Accurate measurement of rice kernel sizes after milling is critical to design, develop and optimize rice milling operations. The size and mass of the individual rice kernels are important parameters typically associated with rice quality attributes, particularly head rice yield. In this study, we propose a novel methodology that combines image processing and machine learning (ML) ensemble to accurately measure the size and mass of several rice kernels simultaneously. We have developed an image processing algorithm with the help of recursive method to identify the individual rice kernels from an image and estimate the size of the kernels based on the pixels a kernel occupies. The number of pixels representing a rice kernel has been used as its digital fingerprint in order to predict its size and mass. We have employed a number of popular machine learning models to build a stacked ensemble model (SEM), which can predict the mass of the individual rice kernels based on the features derived from the pixels of the individual kernels in the image. The prediction accuracy and robustness of our image processing and SEM are quantified using uncertainty quantification. The uncertainty quantification showed 3.6%, 2.5%, and 2.4% for mean errors in estimating the kernel length of small-grain (Calhikari-202), medium-grain (Jupiter), and long-grain (CL153) rice, respectively. Similarly, mean errors associated with predicting the 1000 grain weight are 4.1%, 2.9%, and 4.3% for Calhikari-202, Jupiter, and CL153, respectively. Use of the developed algorithm in rice imaging analyzers could facilitate head rice yield quantifications and promote quicker rice quality appraisals.