基于多源信息融合与深度学习的柚果内部品质无损检测方法研究

To find a nondestructive, fast and low-cost way for pomelo fruit's internal quality (granulation, sugar content, and water content) detection is essential to China's pomelo fruit industry development. Due to the large size and thick skin of pomelo fruit, current technologies grade the pomelo fruit by only detecting the external physical characteristics. Scholars and our previous research results proved the feasibility of detection of pomelo fruit internal qualities by using multiple heterogeneous information (visible/near infrared spectral information, machine vision information and weighing sensor information). However, the key of accurate and stable detection is to guarantee the comprehensiveness and simplification of fusion feature and the learning ability and generalization ability of detection model. Thus, this project propose the study of nondestructive detection of internal quality of pomelo fruit based on multi-source information fusion and deep learning. The project would extract deep characteristics from multiple heterogeneous information of pomelo fruit, reveal the correlated property between the deep characteristics of multiple heterogeneous information and internal quality of pomelo fruit, and build up the analytical model based on deep characteristics of multiple heterogeneous information to describe and explain the internal quality of pomelo fruit. Then, establish data fusion scheme according the analytical model. Finally, the project would build up the detection model of internal quality of pomelo fruit based on multi-source information fusion and deep leaning, which can form a new idea and method for the nondestructive and fast internal quality detection of pomelo fruit. The research results can provide support to the market quality guarantee and scientific postharvest treatment of pomelo fruit, which is significant to promote the pomelo industry development. Also, the research results could provide reference for other large and thick skin fruit's internal quality nondestructive detection.

我国柚果产业亟需一种快速、准确、较低成本的内部品质（粒化、糖度和含水率）无损检测方法，但柚果皮厚果大，已有分级方法仅检测外部特征。相关学者和我们前期研究表明，采用多源融合信息（可见/近红外光谱、机器视觉和称重传感信息）检测柚果内部品质是可行的，保证融合特征提取的全面与精简性以及模型训练的学习与泛化能力是准确、稳定检测的关键。因此，本项目提出基于多源信息融合与深度学习的柚果内部品质无损检测方法研究，从柚果多种异质信息中挖掘深层特征，探明其与内部品质之间的关联特性，进而寻找出适合描述与解释柚果内部品质的多种异质信息深层特征解析模型，以此确定用于建模的柚果多种异质信息深层特征融合方案，结合深度学习的理论和技术提出柚果内部品质检测模型，形成柚果内部品质无损检测新方法与思路。可为保证柚果市场品质、科学采后处理提供支持，对促进柚果产业发展具有重要意义，也可为其他大型厚皮水果内部品质无损检测提供参考。

我国柚果产业亟需一种快速、准确、较低成本的内部品质（粒化、可溶性固形物和含水率）无损检测方法，但柚果皮厚果大，已有分级方法仅检测外部特征，现有方法对内部品质无损检测较困难。本研究以柚果为研究对象，采用多源融合信息（可见/近红外光谱、机器视觉和称重传感信息）与深度学习解决其内部品质无损检测精度不佳的问题。通过研究得到以下几点重要结果：.（1）根据流水线实际分选需求，搭建了柚果光谱、视觉、称重无损检测平台，并设计了适配于柚果检测的传送托盘，参数为：外径100mm、内径80mm、内部夹角17°、聚甲醛树脂且厚度20mm，将该托盘应用在可见/近红外光谱平台采集柚果漫透射光谱数据，具有较好的信噪比；.（2）研究了可见/近红外光谱对柚果硬粒化程度、含水率的识别，结果表明，可见/近红外光谱经SG、GA、MSC、PLSR算法处理后，其柚果硬粒化程度和含水率可被有效的识别，硬粒化程度的测试集无损R2与RMSE可达到0.9738和0.0963，含水率的测试集无损识别R2与RMSE可达到0.7376和0.0489；.（3）研究了柚果可溶性固形物含量与密度、体积、质量的相关特性，发现柚果可溶性固形物含量与密度的关联系最强，且质量越大的柚果体积越大；此外，柚果视觉特征与内部品质指标之间具有较强的非线性相关特性，采用柚果视觉特征对TSS和含水率进行粗略识别是可行的；.（4）研究了采用可见近红外光谱结合卷积神经网络进行深度学习的柚果可溶性固形物含量无损检测精度提升，可将柚果可溶性固形物含量无损检测验证集R2提升0.09，RMSE降低0.23，验证了深度学习可有效挖掘更丰富的特征，提升无损识别精度；.（5）采用可见/近红外光谱特征和机器视觉特征融合结合深度学习算法（SG + GA + CNN + PLSR）相较结合传统识别方法（SG + GA + PLSR）对可溶性固形物含量识别精度训练集R2和RMSE从0.7753和0.4960提升到了0.9636和0.1991，验证集R2和RMSE从0.7714和0.5027提升到了0.8580和0.4276，可有效对柚果可溶性固形物进行识别；.（6）自主研制形成一套流水线式柚果内外品质高精度无损检测分选技术与装备，集成基于多源深度信息融合的柚果品质无损检测技术和装备，在梅州市兴缘农业发展有限公司建立示范基地1个。