Deep learning application for real-time gravity-assisted seed conveying system for watermelon seeds purity sorting

Seed standardization is crucial for all seed breeders as it enables differentiation of performance among specific seed varieties. Standardized seed lots outperform impure seed lots in terms of yield and plant population uniformity. Watermelon growers frequently encounter the challenge of uncertain ploidy seed naming, stemming from the mixture of seedless (triploid) and seeded (tetraploid and diploid) seeds. This uncertainty adversely affects farmers' incomes and hinders the development of specialized watermelon seed enterprises. Watermelon seed purity has been traditionally determined by human expertise methods based on seed thickness, weight, and specific gravity. In this study, we used machine vision and deep learning technology to distinguish triploid (3×) watermelon seeds from diploid (2×) and tetraploid (4×) seeds in real-time. A YOLOv5n deep learning model with over 95 % discrimination of seeded seeds from seedless seeds, was developed and applied to an industrial gravity-feed online sorting system. The deep learning model took 5.4 ms to predict and eject every frame containing seeds in the online system, allowing the system to operate at up to 166 frames per second. With the seed vibration hopper frequency set at a constant magnitude of 35 %, the gravity-feed online system can classify and sort seeds according to their ploidy class, achieving an impressive rate of 14.8 kg/hr. These findings demonstrate the potential of deep learning in automation for real-time seed discrimination and sorting in online systems.