The operating procedures of seed testing are complicated, time-consuming and labor-intensive. In order to improve the efficiency of seed testing, Taiwan Seed Improvement and Propagation Station, Council of Agriculture, Executive Yuan (TSIPS) and the research team of Professor Chen Ching-han of National Central University jointly completed the development of the "digital platform for seed illustration" with artificial intelligence technology in 2021. The system establishes digital processing of seed morphology and the database of seed characteristics for 447 species of plant seeds in 69 families. The system can  rapidly eliminates weed seeds, and effectively  maintain the agricultural ecological security of our country.

TSIPS indicated that Seed Testing Laboratory is a special unit for the government's seed inspection. There are about 1,500-2,000 inspecting samples every year for import and export trade seeds, seeds under multiplication and certification system and seeds of various commercial crops sent by the governments. The results showed that 20-30% samples were mixed with other crop seeds or weed seeds. Once the germinating foreign seeds escape from the field, they will expand rapidly and become invasive plants, which will not only affect the domestic ecology, but also aggravate the costs of field control. Due to the variety of mixed seeds were diverse, inspectors in the past mostly relied on seed specimens and paper illustrations. Not only the inspection was time-consuming, but some of the seed illustrations were old and the pictures were distorted, making it more difficult for inspectors to test.

TSIPS explained that the system uses deep learning and machine learning methods to establish automatic identification functions. Users can upload photos of unknown seeds to the system, and through the function of finding pictures by image, the system will find the 5 most similar seeds within 1 second, and the user will combine the text description of the characteristics of the seeds listed in the system, to quickly identify the correct seed species, thereby simplifying inspection procedures and time. In addition, basic search functions such as family name, scientific name and Chinese name can also be used to search, to meet different requirements of users.

TSIPS said that the system will become a powerful tool for seed gatekeepers, effectively improve the inspection efficiency, and can be used by relevant seed researchers, seed manufacturers and farmers in the future, so that everyone can work together to maintain the agricultural ecological security of our country.

▲Fig1. The digital platform for seed illustration	▲Fig2. The classifier architecture of multimodal features for seed
▲Fig3. Different types of seeds with image processing