SBIR Phase II: Use of Machine Learning Techniques for Robust Crop and Weed Detection in Agricultural Fields

这一小型企业创新研究(SBIR)第二阶段项目旨在进一步开发一种基于计算机视觉的新型植物识别系统，用于农业杂草控制的商业化。该系统将为化学除草剂提供一种具有成本竞争力的替代品，化学除草剂是一个价值200亿美元的全球市场。现有的基于计算机视觉的方法可以将一小块绿色植被从棕色背景中分割出来，但无法提供机械化、自动化除草所需的选择性和精确度。该项目的目标是创建与人眼和大脑的能力相匹配的软件算法，以快速可靠地实时将植物分类为农作物和杂草。该项目团队将建立一个基于分层分类器的计算机视觉算法。该分类器将利用现场定制的支持向量机(SVM)，该支持向量机使用兴趣点而不是基于形状的方法，这是一种新的视觉对象识别方法。这项研究的结果将是创建一种集成到自动除草系统中的算法。该项目的更广泛的影响/商业潜力是巨大的，因为开发一种替代化学密集型农业杂草控制的方法将影响技术理解，创造商业机会，并对社会产生积极影响。在技术上，该项目将通过开发基于兴趣点和支持向量机的实时自动植物识别系统，推动计算机视觉和机器学习领域的发展。在商业上，该系统将为传统农民提供一种有效的、不含化学物质的除草方法，并将为有机农民提供第一种真正精确的有机杂草控制方法。在美国，粮食生产中杂草控制的潜在市场估计为40亿美元。该系统消除化学除草剂使用的能力具有深远的社会影响。美国农民每年仅在玉米和大豆上就使用超过2.5亿磅的除草剂，产生了许多意想不到的有害副作用。河流、湖泊和地下水中的化学物质浓度正在上升，抗除草剂杂草的流行率呈指数级增长。这些化学品的一种替代品在保护环境完整性的同时限制了社会的接触。

This Small Business Innovation Research (SBIR) Phase II project seeks to further develop a novel computer vision based plant identification system for commercialization in agricultural weed control. This system will provide a cost competitive alternative to chemical herbicides, a global $20B market. Existing computer vision based approaches can segment a 'splotch' of green vegetation from a brown background but are unable to provide the selectivity and precision necessary for mechanized, automated weeding. This project's objective is to create software algorithms that match the capability of the human eye and brain to quickly and reliably classify plants into crops and weeds in real-time. The project team will build a computer vision algorithm based on a hierarchical classifier. This classifier will utilize a field customized support vector machine (SVM) that uses point-of-interest rather than shape-based methods, a novel approach to visual object identification. The result of this research will be the creation of an algorithm integrated into an automated weeding system.The broader impact/commercial potential of this project is significant, as the development of an alternative to chemical intensive agricultural weed control will impact technological understanding, create commercial opportunity, and positively impact society. Technologically, the project will advance the fields of computer vision and machine learning through development of a real-time, automated plant identification system based on point-of-interest and SVMs. Commercially, the system will offer conventional farmers an effective and chemical-free method to eliminate weeds, and it will offer organic farmers the first truly precise organic weed control method. The addressable market for weed control in food production is estimated to be $4B in the U.S. The system's ability to eliminate the use of chemical herbicides has a profound societal effect. U.S. farmers apply over 250M pounds of herbicide annually on corn and soybeans alone, with many unintended and detrimental side effects. Chemical concentrations in rivers, lakes and groundwater are rising, and the prevalence of herbicide resistant weeds is growing exponentially. An alternative to these chemicals limits society's exposure while protecting environmental integrity.