Defect Detection and Size Grading of Harvested Japanese Yam Using Computer Vision Technology

Abstract

Root crop yield estimate and harvest technology, especially Japanese Yam harvest production, has a high demand for intelligent automation systems. This paper develops a Japanese Yam quality grading system based on computer vision and deep learning, which can automatically detect the shape and quality of Japanese Yam and grade the size of harvested. Specifically, based on Shuffle-Net and transfer learning, a lightweight deep learning model (CDD Net) was constructed to detect surface and shape defects of Japanese Yam. It's methods were also proposed based on minimum bounding rectangle (MBR) fitting and convex polygon approximation. Experimental results showed that the detection accuracy of the proposed CDD Net was 98.94% for binary classification (normal and defective) and 92.92% for multi-class classification (normal, curve, fork root, fracture), and demonstrated good performance both in time efficiency and detection accuracy. The size accuracy of MBR fitting and convex polygon approximation was 92.8% and 95.1% respectively. This study provides a practical method for defect detection and size grading, which has great potential for application in yield estimation of root crops.