In a study published in Smart Agricultural Technology, a research group led by ZHAO Qinglei from the Changchun Institute of Optics, Fine Mechanics and Physics of the Chinese Academy of Sciences, together with collaborators, has developed a deep learning method for screening high-vigor maize seeds before germination using ordinary RGB images. 

The method, called SeedFocusNet, achieved 74.39% precision in identifying high-vigor seeds. Greenhouse tests showed that the emergence rate of seeds selected by the model was 18–22 percentage points higher than that of unscreened seeds.

Seed vigor reflects a seed's ability to germinate rapidly and develop into a healthy seedling, making it an important indicator for breeding and pre-sowing selection. Conventional vigor tests can be destructive and may require several days of labor. Techniques such as X-ray CT and hyperspectral imaging can provide detailed information, but they rely on specialized equipment. Although ordinary RGB cameras are much cheaper and easier to use, vigor is an internal physiological property, and photographs reveal only subtle differences in seed color, texture, and shape. This makes predicting vigor early on from a single image particularly challenging.

To determine if these subtle visual cues could be used for practical screening, the researchers collected pre-germination RGB images of 955 maize seeds. Then, they conducted controlled germination tests to determine the seeds' actual vigor. They processed the images to remove background interference and extracted informative regions, such as the seed ends, hilum, and possible surface damage. They analyzed these features using SeedFocusNet, which employed Focal Loss to pay more attention to difficult-to-distinguish seeds. Bayesian optimization was then used to automatically search for suitable model parameters, reducing reliance on manual trial and error.

The researchers compared several commonly used deep-learning architectures and found that ResNet50 provided the best overall balance for practical screening. On the independent test set, SeedFocusNet achieved 74.39% precision and 54.46% recall, compared with 67.53% precision and 46.43% recall for an optimized conventional ResNet50 baseline. 

Greenhouse experiments provided further validation. Seeds selected by SeedFocusNet achieved emergence rates of 78%, 76%, and 68% after one, two, and three years of storage, respectively, compared with 60%, 54%, and 48% for unscreened seeds.

This study demonstrates the potential of combining inexpensive RGB imaging with deep learning for early seed-vigor screening. By identifying promising seeds before planting, this approach could provide breeders and seed producers with a more accessible alternative to slower or equipment-intensive testing methods. 

The researchers noted that the current system was evaluated on one maize accession under controlled laboratory conditions, and further testing across different varieties, seed batches, imaging environments, and field conditions will be needed before large-scale practical application.