Multitask Learning for Wheat quality and variety classification

Abstract

Qualitative grading and breed identification of wheat is a challenging and important task. Current processes involve manual grading of wheat which are time-intensive and prone to human errors. Some of the established automation methods which are known to work well utilize statistical approaches to extract physical features that are highly sensitive to camera depth and ambient conditions and have specific hardware requirements. In this paper, we propose a highly robust algorithm that uses digital image processing to eliminate the background and retain the grains in the foreground and optimized neural networks that classify the wheat grains. We have devised two major approaches, specifically tuned to tackle breed and grade classification of wheat grains – a pair of individual models and a multi-task model. The models attain high accuracy, with 92.86% for the multi-task model and 94.01% for the individual model, while also reducing computational time and memory usage. The multi-task model attains a precision of 0.82 and recall of 0.89, whereas the individual model achieves a precision of 0.84 and recall of 0.91. The runtime of the algorithm on the Intel i5 4th gen processor is very low (0.14 s per grain). This is due to the low model size which is approximately 0.93 MB. This solution requires no specific camera resolution or expensive hardware.