【作者】 李志远；

【导师】 罗林保；

【作者基本信息】 合肥工业大学 ， 电子科学与技术， 2024， 硕士

【摘要】 随着社会经济的发展和人民生活水平的提高,人们对水果需求量增大的同时对水果品质的要求也越来越高。加之,我国水果种植面积和产量在世界上首屈一指,但是出口量却很少,在国际市场缺乏竞争力,其主要原因在于水果采后商品化处理技术落后。因此,增强水果的采后产业化、商品化处理,不仅可以满足人们对水果品质的高要求,更是提高我国水果产业竞争力的关键。可溶性固形物含量(soluble solids content,SSC)是表征水果内在品质的一个重要参数。与复杂的化学方法相比,光谱分析具有快速、无损等优点,已被广泛用于水果SSC的建模。然而,由于从单一光谱范围获得的信息不足,模型的性能受到限制。探索多传感器光谱融合方法以寻求更好的性能是非常有价值的。在本学位论文中,以产自我国的砀山梨为研究对象,以其SSC为检测指标,利用近红外光谱、高光谱成像和光谱融合技术,结合化学计量学方法,对砀山梨SSC的检测进行了建模分析,并基于现场可编程门阵列(Field Programmable Gate Array,FPGA)设计与实现了改进的一维卷积神经网络(one-dimensional convolutional neural network,1D CNN),完成了对砀山梨SSC的检测。首先,研究了近红外光谱、高光谱成像技术的原理与特点以及其在水果SSC检测领域的应用,为本研究提供了理论基础。其次,设计了砀山梨样本近红外光谱数据和高光谱数的采集实验,并建立砀山梨光谱数据库。针对光谱数据存在异常值和噪声干扰等问题,对数据进行光谱预处理,以提高数据集的质量和可靠性。利用多传感器光谱融合技术,针对砀山梨SSC定量检测问题,建立了不同的回归预测模型。将光谱融合前后的数据分别建模,验证多传感器光谱融合策略的优越性。对1D CNN进行了改进,将全局上下文(Global Context,GC)模块引入1D CNN中,从通道维度和空间维度提升网络性能。通过与1D CNN和传统回归模型的比较,验证了1D GC-CNN的有效性和鲁棒性,其最佳检测结果R_p~2、RMSEP和MAEP分别是0.9012、0.2788和0.2220。最后,对算法进行了FPGA设计与实现,分析研究了算法运算时可能存在的并行性,对本文的砀山梨SSC检测算法模型各层的计算量进行了评估。在ZYNQ7020上实现了砀山梨SSC检测系统,增强了其实时性和适用性,为工业应用提供了快速和可扩展等方面的优势。这一进展不仅为水果的光学特性建模提供了一种新的有效方法,还为其他农产品的无损分析开辟了途径。更多还原

【Abstract】 With the development of social economy and the improvement of people’s living standard,people’s demand for fruits increases while the requirement for fruit quality is also higher and higher.Although China ranks first in the world in terms of fruit cultivation area and production,its export volume is quite low,lacking competitiveness in the international market.The main reason for this is the backward post-harvest commercialization technology.Therefore,enhancing the post-harvest industrialization and commercialization of fruits is not only crucial to meet the high-quality demands of consumers but also essential to improve the competitiveness of China’s fruit industry.Soluble solids content(SSC)is an important parameter to characterise the intrinsic quality of fruits.Compared with complex chemical methods,spectral analysis has the advantages of being fast and non-destructive,and has been widely used for modelling fruit SSC.However,the performance of the models is limited by the insufficient information obtained from a single spectral range.It is valuable to explore multi-sensor spectral fusion methods in search of better performance.In this dissertation,Dangshan pear,which is produced in China,is taken as the research object,and its SSC is used as the detection index,and near-infrared spectroscopy,hyperspectral imaging and spectral fusion techniques,combined with chemometrics methods,are used to carry out modelling and analysis of the detection of SSC of Dangshan pears,and an improved one-dimensional convolutional neural network(1D CNN)based on Field Programmable Gate Array(FPGA)has been designed and implemented to complete the detection of the SSC of Dangshan pears.Firstly,the principles and characteristics of near-infrared spectroscopy and hyperspectral imaging techniques and their applications in the field of fruit SSC detection were investigated to provide a theoretical basis for this study.Secondly,the experiments of collecting NIR spectral data and hyperspectral numbers of Dangshan pear samples were designed,and the Dangshan pear spectral database was established.Aiming at the problems of outliers and noise interference in the spectral data,spectral pre-processing of the data was carried out to improve the quality and reliability of the dataset.Using multi-sensor spectral fusion technology,different regression prediction models were established for the problem of quantitative detection of SSC of Dangshan pear.The data before and after spectral fusion are modelled separately to verify the superiority of the multi-sensor spectral fusion strategy.The one-dimensional convolutional neural network is improved by introducing the Global Context(GC)block into the 1D CNN to enhance the network performance in terms of channel dimension and spatial dimension.The effectiveness and robustness of 1D GC-CNN is verified by comparing with 1D CNN and traditional regression models,and its best detection results R_p~2,RMSEP,and MAEP are 0.9012,0.2788,and 0.2220,respectively.Finally,the FPGA design and implementation of the algorithm was carried out to analyse and study the possible parallelism in the algorithm operation,and the computational volume of each layer of the Dangshan pear SSC detection algorithm model in this paper was evaluated.The implementation of the Dangshan pear SSC inspection system on the ZYNQ7020 development board enhances its real-time and applicability,providing advantages such as speed and scalability for industrial applications.This advancement not only provides a new and effective method for modelling optical properties of fruits,but also opens up avenues for non-destructive analysis of other agricultural products.更多还原