OpenSeed: Towards an Open-Source Seed Image Analysis Ecosystem

1. Accurate seed identification is crucial for seed quality evaluation, germplasm conservation and related applications. Seed image analysis has developed rapidly in recent decades but remains fragmented and confined to small, closed-source datasets and outdated models. 2. We created OpenSeed-LZU, a generic, large-scale, open-source dataset consisting of over 300,000 seed images from 656 species, which can be used in many downstream tasks like seed detection or building multimodal seed large language models. 3. We trained and evaluated 14 advanced deep learning models spanning three computational power levels and achieved species-level identification accuracy of 90.67% (scratch-trained) and 93.05% (pre-trained). We further preliminarily explained the models using Class Activation Maps. The inference throughputs were benchmarked across five hardware devices to guide deployment. 4. OpenSeed, a user-friendly, cross-platform application that facilitates rapid seed identification for research and education is developed. This study aims to establish an open ecosystem for seed image analysis to accelerate the application of artificial intelligence in seed science.