结合稀疏表示和深度学习的伏牛山猕猴桃表面缺陷动态检测分类模型研究

Funiu Mountian kiwi fruit is a main economic crop in the water source area of the middle route project of the south-to-north water transfer. The industrialization of this kind of fruit is of great significance demonstration to improve the local economic development. At this stage, due to the lack of a new reliable and systematic theoretical model, the postharvest automatic grading technology based on fruit surface defect detection and classification can not be effectively implemented. The rough surfaces of kiwi fruits increase the difficulty to defect detection, and the diversity of defect types increases the complexity of defect classification. In recent years, the theories of sparse representation and deep learning have made breakthrough progresses in the fields of image denoising and image classification respectively. Given this, the purpose of this project is to carry out research of Funiu Mountain kiwi fruit surface defect dynamic detection and classification model based on sparse representation and deep learning. It has important scientific significance to systematically improve the agricultural non-destructive grading theoretical level based on computer vision. The main research strand of this project is to construct a complete dynamic fruit surface defect detection and classification model, which includes the fruit image denoising, surface defect detection, surface defect classification and surface defect statistics. The main research topics include: Firstly, the construction of surface defect detection model based on multi-core dictionary sparse filter and color space orthogonal transformation will be developed. Secondly, we will focus on the surface defect classification theory and technology combined with deep learning and sparse representation. Finally, we will propose a multi-perspective and multi-feature surface defect tracking and quantitative analysis theory.

伏牛山猕猴桃作为南水北调中线水源地的特色经济作物，对带动水源地的农牧业发展具有重大的示范意义。现阶段由于缺少可靠系统的新理论模型，基于猕猴桃的表面缺陷分级技术无法有效的实施。猕猴桃粗糙的果皮给缺陷检测增加了难度，缺陷种类的多样性又增大了分类的复杂性，而近年来稀疏表示和深度学习理论分别在图像去噪和分类的部分领域取得了突破性进展。鉴于此，本项目拟开展结合稀疏表示和深度学习的伏牛山猕猴桃表面缺陷动态检测分类模型的研究，这对于系统化提高基于计算机视觉的农产品无损分级技术的理论层次具有重要的科学意义。本项目将以猕猴桃表面缺陷的去噪检测、分类及跟踪统计为主线构建完整的动态检测分类模型，主要研究内容包括：基于多核字典稀疏滤波和颜色空间正交变换的表面缺陷去噪检测模型的构建；结合深度学习和稀疏表示的表面缺陷分类理论和技术；多视角多特征表面缺陷跟踪与定量分析理论与方法。

本项目利用计算机视觉和人工智能理论研究伏牛山猕猴桃表面缺陷动态检测分类模型的构建。其一，本项目的研究对象伏牛山猕猴桃的表皮粗糙，通常覆盖着一层细细的绒毛，因此不同于苹果、芒果等表皮光滑的水果，为了能够提高检测精度，在对缺陷检测时表面噪声的去除就显得尤为重要。其二，本项目是以工厂流水线环境中猕猴桃表面缺陷的动态检测为背景，因此所构建的多视角下的动态缺陷检测模型应具有一种表面缺陷统计分析流程。以此为研究对象，针对这些特点，我们主要从表面缺陷的滤波去噪、特征分类和运动跟踪三个方面进行了相关通用理论算法的研究，取得了系统的研究成果，并推广到了多个应用领域，主要包括：1）深入研究了基于稀疏表示的通用滤波器设计及一种基于维度变换的降维空间下的滤波器框架设计。2）针对动态环境下的缺陷统计，研究了在粒子滤波框架下基于多种特征融合的视觉跟踪算法及相应的光源设计问题。3）研究了基于缺陷分类和深度图重建的深度学习网络，尝试融合表面信息及深度信息来增强缺陷的跟踪和分类。4）利用项目的开展有力推动了师范类地方高校在应用型转型发展中的工科学科建设并取得了突破性进展。项目按照计划书要求执行，完成了计划书所列研究内容，达到预期指标。其研究成果为下一步产学研推广奠定了理论依据和方法基础。