Identification of broken rice rate based on grading and morphological classification

The broken rice rate (BR) is a critical metric which influences the appearance, processing, and economic value of rice. However, current machine vision and machine learning approaches engender significant errors when calculating BR. This study introduces a novel restoring method for identifying BR by leveraging grading and morphological features. A three-class classification model using Convolutional Neural Network (CNN) was devised to distinguish broken rice types of crescent head, elliptical tail, and quadrilateral midst based on their morphological characteristics. After training, the accuracy of classfication model is over 98.7%. Taking the longest 10% of rice grains in the image to be identified as head rice references, the broken grains are filtered by calculating the length proportion to the head rice via machine vision. The filtered broken grains are classified to one of three morphological categories with the trained CNN. The broken grains are virtually 'restored' to head rice equivalents based on the classified shape and the grading size. Finally, the BR is determined by comparing the counts of original and restored grains. The results of two testing conditions which including all and lacking some broken grains demonstrate that the proposed method can accurately and effectively identify the BR in real-time (2.5s). • Virtually restoring the broken rice grains to head rice equivalents • Classifying broken grains to seven categories based on shapes and ratios • Mitigating errors arised by samples, acquiring devices and focal lengths • Simplifying the seven-class classification into a three-class problem • Accurately identifying broken rice rate in real-time