New edge-powered vision model transforms safe, high-throughput screening of Aspergillus-infected seeds

Built on lightweight deep learning architectures and optimized for embedded devices, the system achieves precise segmentation of infected regions, accurate separation of adjacent seeds, and consistent calculation of infection indices.

Peanut, maize, and rice are global staple and oilseed crops, yet all are vulnerable to contamination by A. flavus, a fungus that produces carcinogenic and mutagenic aflatoxins. Breeders urgently require accurate assessments of infection levels to develop resistant cultivars, but current manual scoring approaches expose personnel to hazardous spores and suffer from inconsistent results. At the same time, conventional computer vision systems often demand high computing power, making real-time field deployment difficult. Edge computing—processing data locally on compact, low-power hardware—offers a promising pathway for bringing intelligent fungal detection directly to farms and breeding facilities. Based on these challenges, there is a critical need for a fast, safe, and transferable on-site evaluation system.

A study (DOI: 10.1016/j.plaphe.2025.100110) published in Plant Phenomics on 18 September 2025 by Yande Liu’s & Dapeng Ye’s team, Xiamen University of Technology & Fujian Agriculture and Forestry University, provides a safe, scalable, and automated solution for breeding programs and post-harvest quality inspection.

In this study, the researchers employed an enhanced edge computing-based computer vision framework (Edge CV) to evaluate fungal infection in crop seeds, integrating three major methodological components: an improved segmentation model, a refined post-processing pipeline, and transfer learning for cross-species adaptability. The segmentation model incorporated a Convolutional Block Attention Module (CBAM) to regulate channel–spatial information flow and optimize feature extraction, supported by joint losses for bounding boxes, classes, and pixel-wise segmentation. Post-processing techniques—including morphological operations, connected components analysis, and a watershed algorithm—were applied to merge mold and unmold regions belonging to each seed and to accurately separate adjacent seeds. Finally, deep transfer learning was used to adapt the system to maize and rice, enabling evaluation beyond peanut datasets. Corresponding to these methods, the segmentation results demonstrated strong computational efficiency and accuracy, with the optimized model achieving 89.7% accuracy and high inference speed. Training curves stabilized after 100 epochs, and precision and recall exceeded 97% for both bounding boxes and masks, with mAP50 values reaching 97.6–97.9%. Ablation experiments further confirmed that both the CBAM module and the post-processing pipeline independently improved performance, and together raised mAP50:95 to 89.7% without increasing inference time, as post-processing required only 0.5–2 ms per seed. Post-processing effectively merged infected and healthy regions of individual seeds and accurately separated closely connected ones, enabling reliable infection grading. When applied to on-site peanut evaluations, Edge CV produced highly consistent infection indices (±0.01% fluctuation) and showed close agreement with manual measurements (R² = 0.991, RMSE = 0.007) while reducing evaluation time by three orders of magnitude. Finally, transfer learning enabled successful application to maize (R² = 0.968) and rice (R² = 0.949), demonstrating robust adaptability across seed types despite challenges associated with illumination variability and seed occlusion.

Edge CV offers a transformative tool for agricultural breeding, seed inspection, and food safety monitoring. By eliminating direct human contact with A. flavus, the system greatly enhances safety during fungal assessment. Its high throughput and consistent accuracy allow breeders to rapidly evaluate resistance levels across large seed populations, accelerating selection programs. For industry, the technology enables real-time, on-site screening during post-harvest handling, potentially reducing aflatoxin contamination along the supply chain. Its lightweight architecture and transferability also make it suitable for integration into agricultural robotics, automated sorting equipment, and mobile diagnostic platforms.

###

References

DOI

10.1016/j.plaphe.2025.100110

Original Source URl

https://doi.org/10.1016/j.plaphe.2025.100110

Funding information

This work was supported by the National Key R&D Program Project, Research and Development of Intelligent and Efficient Processing Technology and Equipment for Vegetable Production Areas(2023YFD2001301). The authors thank the China Scholarship Council (CSC No. 202408350068) for the financial support to the author (Libin Wu) to conduct her doctoral research in the Department of Bioresource Engineering at McGill University.

About Plant Phenomics

Plant Phenomics is dedicated to publishing novel research that will advance all aspects of plant phenotyping from the cell to the plant population levels using innovative combinations of sensor systems and data analytics. Plant Phenomics aims also to connect phenomics to other science domains, such as genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics, and computer sciences. Plant Phenomics should thus contribute to advance plant sciences and agriculture/forestry/horticulture by addressing key scientific challenges in the area of plant phenomics.