【作者】 张若宇；

【导师】 应义斌；

【作者基本信息】 浙江大学 ， 农业电气化与自动化， 2014， 博士

【摘要】 果蔬品质无损检测与自动化分级是农产品加工领域的热点课题,对满足消费者日益增高的食品品质和安全要求,提高果蔬市场竞争力,增加农民收入都具有重大意义。虽然我国是果蔬生产大国,但由于产后加工和处理水平不高,因而在国际市场上缺乏竞争力。与形状、尺寸、颜色和缺陷等外部品质参数一样,化学成分和质地也是评价果蔬品质的重要指标,决定着产品的品质。然而这些参数指标无法直接通过视觉观察,因此其检测难度较大。本研究以不同产地的鲜食和加工番茄为研究对象,选取代表番茄化学成分和质地的品质参数可溶性固形物含量(SSC)和硬度(包括表皮硬度F1、果肉硬度F2和修正硬度Fxz)为检测指标,搭建反射、散射和漫透射三种高光谱成像系统,获取番茄空域光谱和纹理参数(或拟合参数),利用偏最小二乘回归(PLSR)方法建立番茄SSC和硬度的定量分析模型。探索在三种高光谱成像模式下番茄SSC和硬度的检测方法,为进一步研发番茄化学成分和质地的综合品质参数在线检测装备提供理论依据。主要研究结果和结论如下：(1)提出了基于主成分降维思想的背景分割方法,通过对样本的主成分图像进行背景分割进而获取目标对象的光谱及纹理信息。该方法较单波段分割法可以更加有效地去除图像背景。(2)在高光谱反射和漫透射两种成像模式下,对两种番茄光谱图像像素点光谱校正方法进行了考察。结果表明,高光谱反射成像模式下,两种校正方法对SSC的检测效果均有所改善,并且像素点光谱归一化校正效果相对较好(模型RPD=2.315)；高光谱漫透射成像模式下,像素点光谱归一化校正方法同样对SSC的检测效果有所改善(模型RPD=1.983),而像素点光谱标准正态变量校正后的模型鲁棒性差于校正前。总体而言,像素点光谱归一化校正在两种成像模式下,均能够提高番茄SSC的检测效果,具有更好的适应性。(3)对比分析了高光谱反射成像和近红外漫反射光谱技术对番茄SSC和硬度检测效果。结果表明,采用高光谱反射成像技术获取番茄整体空间光谱信息,对番茄SSC和硬度进行检测优势明显。其中采用高光谱反射成像检测番茄SSC、 F1、 F2和Fxz模型的RPD分别为2.071,1.150,1.150,1.345。(4)提出了新的高光谱成像模式,并同时开展了三种高光谱成像模式下对番茄SSC硬度的检测研究。研究发现：1)三种成像模式下平均光谱相对优于纹理参数(或拟合参数)对番茄SSC和硬度测效果。其中,番茄SSC的检测效果最好,而且整体平均光谱对该指标的检测模型的RP1均接近或大于2。硬度检测中,Fxz的检测效果最好。但相对于SSC而言,光谱成像对度的检测效果不十分理想,模型的RPD均小于1.5。2)对单一高光谱成像模式而言,高光谱漫透射成像可能由于可以获取番茄深层光信息,因此在对番茄SSC检测中较高光谱反射和高光谱散射成像具有一定优势。由于高光谱反射成像可以获取番茄较大像素区域的光谱信息,在对Fxz检测中比高光谱散射和漫透射成像技术更具优势。并且在番茄SSC和Fxz的检测中,通过多个姿态获取番茄整体平均光谱的检测方法整体优于局部小像素区域平均光谱的检测方法。3)从不同成像模式光谱信息组合对番茄SSC和硬度检测效果来看,不同成像模式光谱信息组合对番茄SSC和硬度的检测效果较单一成像模式来说无明显的优势,在硬度检测中,依然是Fxz检测效果最好,F1和F2检测效果较差。因此,今后在番茄SSC和硬度检测中,可以考虑采用Fxz来表征番茄的硬度特性,然后利用高光谱漫透射成像技术检测番茄SSC,利用高光谱反射成像技术检测番茄Fxz,以实现番茄SSC和硬度的准确检测。更多还原

【Abstract】 Nondestructive quality inspection of fruits and vegetables as well as automatic grading is a hot topic in the agro-processing research area. It is of great importance to satisfy consumers" increasing demands for food quality and safety, to ensure the quality of fruits and vegetables and enhance their market competitiveness and to improve farmers’income. Although there are a great amount of fruits and vegetables in our country, the market competition capability is still left behind in the international market because of insufficient and low-tech processing of post-harvest. Internal ingredients and characters, along with external quality indicators including shape, size, color, and defects are important indexes to determine the quality of products. However, it is a difficult task to assess the indicators, because they cannot be visualized directly. Tomatoes from different regions were chosen as the study objects of this thesis. Soluble solid content (SSC) and firmness including peel firmness Fl, flesh firmness F2and corrected finnness Fxz were detected to represent internal ingredients and characters of tomatoes. Hyperspectral imaging systems in reflectance, scattering and diffuse transmittance modes were constructed and used as platforms to obtain spatial spectra and texture parameters (or fitting parameters). Then quantitative prediction models of SSC and firmness were established by using partial least square regression (PLSR). With the help of hyperspectral imaging systems in reflectance, scattering and diffuse transmittance modes, potential applications for detecting SSC and firmness were investigated to establish a theoretical basis for online detection of ingredients and characters simultaneously. The main results are as follows:(1) A background segmentation method based on principle component score images was proposed, and was utilized to obtain mean spectra and texture information of target objects. This method is more effective in wiping out background compared with single-band segmentation.(2) The methods for spectral correcting of pixels were applied in tomato spectral images under hyperspectral reflectance and diffuse transmittance modes. Results showed that these two correction methods could improve the detection of SSC, and the pixel spectral normalization was better (RPD=2.315) under reflectance mode; and could also enhance results which were obtained after spectral nonnalization of pixels under diffuse transmittance mode (RPD=1.983), while the robustness was slightly worse after correction using pixel spectral standard norm*variate. Overall, the pixel spectral normalization can advance the detection of SSC unde hyperspectral reflectance and diffuse transmittance modes.(3) Comparative analysis of prediction results based on hyperspectral reflectance imaginj and near-infrared diffuse transmission spectroscopy showed that hyperspectral reflectanc imaging performed much better in predicting SSC and firmness of tomatoes by obtaining overal spatial spectra information in detecting SSC, F1, F2and Fxz, with RPD=2.071,1.150,1.1501.345, respectively.(4) Hyperspectral diffuse transmittance imaging was proposed as a new hyperspectra imaging mode, and SSC and firmness values were measureed by three hyperspectral imaging modes. Results showed as following:1) Mean spectra performed better than textural parameters (or fitting parameters) in detectioi of SSC and firmness under the three hyperspectral imaging modes, in which the detection o SSC was the best with RPD being close to2or larger than2. The detection of Fxz was the bes in firmness detection, however, compared with SSC detection resultes, spectral imaging foi firmness testing is not very satisfactory with the RPD<1.5.2) For single hyperspectral imaging modes, the hyperspectral diffuse transmittance imaging has certain advantages in predicting SSC compared with reflectance and scattering imaging because of its accessibility to spectral information in deeper tissue of tomatoes. Hyperspectra reflectance imaging was superiority in detecting Fxz compared with hyperspectral scattering anc diffuse transmittance imaging mode, because spectral information in bigger pixel areas was utilized. Overall, mean spectra of tomato obtained by combined positions was more suitable foi detecting SSC and Fxz than part pixels areas of tomato.3) The results of predicting SSC and firmness using combined spectra information, which obtained by different imaging modes, showed no obvious advantage compared with the single imaging modes. The Fxz performed best in detecting firmness, while F1and F2detection showed poor results.Therefore, Fxz was considered as the best parameter to characterize the firmness of tomatoes And hyperspectral diffuse transmittance imaging is more suitable for detecting SSC, while hyperspectral reflectance imaging is more suitable for Fxz detection.更多还原