基于计算机视觉的玉米种子表型特征提取研究综述

Summary of Research on Extraction of Phenotypic Features of Maize Seeds Based on Computer Vision

摘要: 玉米种子表型特征是评估玉米产量和品质的关键指标。传统的人工考种方法存在耗时、主观性强及破坏性等局限。本文系统阐述计算机视觉技术在果穗与籽粒两个维度的研究进展，具体分析果穗维度的几何参数测算、秃尖量化与穗行数提取，以及籽粒维度的几何参数测算与内部品质无损检测等核心技术。同时系统梳理了从传统图像处理、经典机器学习、有监督深度学习，再到自监督大模型的技术演进历程。总结开源数据集匮乏、跨场景泛化衰减与边缘部署受限等挑战，并展望视觉大模型与多模态数字融合的发展趋势，旨在为智能考种装备的研发以及玉米精准育种提供坚实的理论参考。

Abstract: The phenotypic characteristics of maize seeds are the key indexes to evaluate the yield and quality of maize. The traditional manual seed testing method is time-consuming, subjective and destructive. In this paper, the research progress of computer vision technology in ear and grain dimensions is systematically expounded, and the core technologies such as geometric parameter calculation of ear dimension, bald tip quantification and ear row number extraction, geometric parameter calculation of grain dimension and nondestructive testing of internal quality are analyzed in detail. At the same time, it systematically combs the technological evolution process from traditional image processing, classic machine learning, supervised deep learning, and then to self-monitoring big model. The challenges such as lack of open source data sets, cross-scene generalization attenuation and limited edge deployment are summarized, and the development trend of visual large model and multimodal digital fusion is prospected, aiming at providing a solid theoretical reference for the research and development of intelligent seed testing equipment and corn precision breeding.