【作者】 张宇；

【导师】 宋平； 郑维刚；

【作者基本信息】 沈阳农业大学 ， 电子信息硕士（专业学位）， 2025， 硕士

【摘要】 文冠果是一种兼具生态效益与经济价值的木本油料植物,其种子品质的核心评价指标为含油含水率。为了更好的判断文冠果的品质,本文基于低场核磁共振（low field nuclear magnetic resonance,LF-NMR）技术结合语义分割模型对比得出针对文冠果含油含水率检测及判别的最优方式,以期能够快速而准确的评估文冠果种子的品质特征。本文的主要研究内容及研究结果如下:（1）基于LF-NMR技术的文冠果含油含水率不同检测方法研究。针对LF-NMR技术检测文冠果含油含水率会出现水油信号重叠,严重干扰检测准确性和可靠性的难题,本文以国标法作为基准进行对比试验,基于LF-NMR的技术特性,提出两种解决方式。一种是一维核磁共振波谱及质子密度加权像结合干燥去水的传统LF-NMR检测方式。另一种是基于纵向弛豫时间（longitudinal relaxation time,T₁）及横向弛豫时间（transverse relaxation time,T₂）的T₁-T₂加权法检测方式。研究结果表明:质子密度加权像显示文冠果在干燥过程中会产生影响LF-NMR检测的物质。从其波谱分析发现,干燥使文冠果自由水与油脂混合峰的总自由度增加,同时氢质子含量上升。因此,传统的LF-NMR技术不利于检测文冠果含油含水率。T₁-T₂加权法与文冠果的含油含水率建立的校准曲线展示良好的线性相关,其决定系数分别为0.9924、0.9957,表明该方法能够准确的判断文冠果的含油含水率。此外,T₁-T₂加权法解释了传统的LF-NMR技术检测过程中存在的问题,文冠果富含不饱和脂肪酸,加热过程中易发生反应,产生同样含有氢元素的有害物质,从而干扰该检测方法的可靠性。因此,基于T₁-T₂加权的LF-NMR技术是检测文冠果含油含水率的最佳方式。（2）基于T₁-T₂加权像的文冠果含油含水率高低判别模型研究。采集文冠果T₁-T₂加权像建立数据集,构建DeepLabV3+、UNet及PSPNet判别模型,对比3种语义分割模型的实际分割效果选择最优模型并进行优化处理。研究结果表明:从模型效能来看,DeepLabV3+平均交并比及平均像素精度均优于UNet及PSPNet;从计算成本来看,UNet所用的模型参数量及计算量最小;从模型可视化来看,裂痕对于文冠果高油以及低油的分割效果具有重要的影响,DeepLabV3+分割的效果更好。综合分析,DeepLabV3+在文冠果含油含水率的分割效果最出色。针对该模型在分割时参数量较大的问题进行轻量化改进,选取Mobilenet V3网络对其主干网络进行重构,并将压缩和激励注意力模块替换高效的通道注意力机制模块。最后对空洞空间金字塔池化模块进行结构以及卷积模块的调整,在减少参数量与计算量的同时减少了轻量化网络带来的精度损失问题。改进后的EM-DeepLabV3+-L网络模型参数量相比原有模型降低了75.77%,计算量下降了71.04%,平均分割速度时间降低了48.27%,同时平均交并比和像素准确率也小幅度上升。因此,EM-DeepLabV3+-L是判别文冠果含油含水率高低的最佳模型。更多还原

【Abstract】 Xanthoceras sorbifolium Bunge（X.sorbifolia）is a woody oilseed plant with both ecological benefits and economic value,and the core evaluation index of its seed quality is the oil and moisture content.In order to better judge the quality of X.sorbifolia,this paper presents the optimal way to detect and discriminate the oil and moisture content of X.sorbifolia based on the low field nuclear magnetic resonance（LF-NMR）technology combined with the semantic segmentation model,so as to quickly and accurately evaluate the quality of X.sorbifolia seeds.The main research content and results of this paper are summarized as follows:（1）Research on different detection methods of oil and moisture content of X.sorbifolia based on LF-NMR technology.For the LF-NMR technology to detect the oil and moisture content of X.sorbifolia there will be water and oil signal overlap,which seriously interferes with the accuracy and reliability of the detection of the difficult problem,this paper uses the national standard method as a benchmark for comparison tests to propose two solutions.One is the one-dimensional nuclear magnetic resonance spectroscopy and proton density-weighted image combined with drying and dewatering of the traditional LF-NMR detection method.The other is a T₁-T₂ weighted method based on longitudinal relaxation time（T₁）and transverse relaxation time（T₂）.The results showed that the Lab value of the proton density weighted image of X.sorbifolia showed an increasing trend after 4 h of drying,which was not consistent with the experimental results of the national standard method,suggesting that substances affecting the changes of LF-NMR detection were generated inside X.sorbifolia.From the spectral analysis,it was found that the drying of X.sorbifolia increased the total degrees of freedom of the free water and oil mixture peaks,while the hydrogen proton content rose.Therefore,the conventional LF-NMR technique is not suitable for the detection of oil and moisture content of X.sorbifolia,and the calibration curves established by the T₁-T₂ weighted method and the oil and moisture content of X.sorbifolia demonstrated good linear correlation with the coefficients of determination of 0.9924 and 0.9957,which indicated that the method was able to accurately determine the oil and moisture content of X.sorbifolia.In addition,the T₁-T₂weighted method explains the problems in the detection process of the traditional LF-NMR technique,in which the X.sorbifolia is rich in unsaturated fatty acids,which are prone to react during the heating process and produce harmful substances also containing hydrogen elements,thus interfering with the reliability of this detection method.Therefore,the LF-NMR technique based on T₁-T₂weighting is the best way to detect the oil and moisture content of X.sorbifolia.（2）Research on the discrimination model of oil and moisture content of X.sorbifolia based on T₁-T₂ weighted images.A dataset was collected from T₁-T₂ weighted images of X.sorbifolia,and DeepLabV3+,UNet and PSPNet discriminative models were constructed,and the optimal model was selected by comparing the actual segmentation effects of the three semantic segmentation models and optimized.The results showed that for both types of datasets,the three models showed consistent segmentation effects.In terms of model performance,DeepLabV3+has better mean intersection over union and mean pixel accuracy than UNet and PSPNet;in terms of computational cost,UNet has the smallest number of model parameters and the smallest number of floating point of operations per second;and in terms of model visualization,cracks has an important effect on the segmentation effect of high and low oil in X.sorbifolia,and the segmentation effect of DeepLabV3+is better.From the visualization of the model,cracks have an important effect on the segmentation effect of X.sorbifolia with high oil and low oil,and DeepLabV3+has a better segmentation effect.In a comprehensive analysis,DeepLabV3+is the most effective in segmenting the oil and moisture content of X.sorbifolia.Lightweight improvement for the problem of large number of parameters in the segmentation of this model,the Mobilenet V3 network was selected to reconfigure the backbone network,and the squeeze and excitation attention module was replaced with the efficient channel attention module.Finally,the structure of the atrous spatial pyramid pooling module was adjusted to reduce the number of parameters and floating point of operations per second while reducing the loss of accuracy caused by the lightweight network.The improved EM-DeepLabV3+-L network model reduces the number of parameters by 75.77%,the floating point of operations per second by 71.04%,and the average segmentation speed time by 48.27%compared with the original model,while the mean intersection over union and mean pixel accuracy also increased slightly.更多还原