Authors

Abstract

Addressing the critical window for weed control in corn at the 2~5-leaf stage—where seedlings and weeds are morphologically similar and highly intermixed in complex field environments—and addressing shortcomings of traditional detection models such as rigid architecture, redundant features, and the difficulty of balancing accuracy with real-time performance, this study proposes an intelligent corn seedling-weed detection model based on Efficient Neural Architecture Search (ENAS), using corn seedlings and field weeds at the 2~5-leaf stage in Heilongjiang Province as test subjects.

Abstract in Chinese

针对玉米2~5叶除草关键窗口期，幼苗与杂草形态相近、田间复杂环境下混杂度高，且传统检测模型结构固定、特征冗余、精度与实时性难以兼顾的问题，以黑龙江省2~5叶期玉米幼苗及田间杂草为对象，提出了一种基于高效神经网络架构搜索（ENAS）的玉米苗草智能检测模型。以YOLOv8为基线网络，设计协同优化搜索策略，对Backbone、Neck与Head均实施细粒度结构搜索：在Backbon层面，通过细粒度搜索强化对玉米2-5叶期幼苗与杂草细微形态差异的特征提取能力；在Neck层面，开展细粒度特征融合结构搜索，在保证多尺度特征充分交互与融合的同时降低结构冗余；在Head层面，进行细粒度检测头结构搜索，优化目标检测的输出适配性；同步引入损失函数自适应搜索机制，实现模型结构与训练目标的联合优化。