Diagnosing Grass Seed Infestation: Convolutional Neural Network-Based Terahertz Imaging

Grass seed infestation is a significant issue in the Australian sheep industry. Detecting the seeds when they are in wool or on the surface of the skin could assist with prevention of the grass seed infestation. Terahertz imaging provides a viable option for detecting seeds due to its short wavelength, non-ionizing feature, and penetration ability through wool. Here we demonstrate that accuracy of seeds detection can be improved utilising a Convolutional Neural Network even when the seeds are not visually distinguishable in terahertz images. Our studies reveal accuracies of greater than 95% and 67% can be achieved in identification of seed hidden underneath 1 cm and 2 cm thick wool under normal incidence. Moreover, our analysis finds that terahertz frequencies in the 0.3–0.4 THz range have better overall classification accuracy compared to other frequency bands. The combination of machine learning and terahertz imaging has the potential to be widely implemented in rapid and on-site detection of grass seed infestation with high efficiency.