Location: Tropical Crops and Germplasm Research
Project Number: 6090-21000-063-012-S
Project Type: Non-Assistance Cooperative Agreement
Start Date: Jun 15, 2026
End Date: May 30, 2027
Objective:
Common bean (Phaseolus vulgaris) is the most important pulse crop globally yet breeding and seed evaluation pipelines remain constrained by slow, labor-intensive, and subjective post-harvest measurements of yield components and seed quality traits. Manual seed counting and laboratory-based imaging workflows limit throughput and delay selection decisions, particularly when evaluating seed number, size, mass, color, and seed damage traits across large breeding populations. This project proposes the development of an AI-enabled, stand-alone edge vision system augmented with a language-model–based descriptive reasoning pipeline for real-time common bean seed yield and quality assessment.
Approach:
Deep learning–based computer vision models will autonomously quantify seed number and extract size, shape, color, and damage-related features. Initial hyperspectral imaging will be used to identify informative spectral bands associated with seed quality and robustness to environmental variability, enabling the design of a cost-effective smart camera system. To enhance usability and scientific interpretation, extracted quantitative features will be passed to a lightweight large language model (LLM)–based reasoning module that generates standardized, human-readable descriptions of seed samples (e.g., “high proportion of uniform, market-acceptable seeds with minimal discoloration”). Model training and validation will leverage SCINet high-performance computing resources.