【摘要】 针对传统果蔬品质检测方法中因样本数量不足而导致检测误差大的问题,提出了一种基于面光源下光子传输模拟的苹果品质检测方法。以苹果为研究对象,采用蒙特卡洛方法仿真光子在苹果双层平板模型的运动轨迹,快速得到20 000幅苹果组织表面光亮度分布图像,以光学参数作为标签,输入卷积神经网络进行训练,将得到的模型进行微调迁移,应用到少量实测苹果光谱图像的数据集上进行光学特性参数的反演,最后将该网络模型全连接层的输出结果与苹果品质建立关联,实现对苹果糖度及硬度的无损检测。结果表明,果肉吸收系数μ_a2反演准确率为93.24%,果肉散射系数μ_s2反演准确率为92.54%;与传统光学参数方法相比,苹果品质分类模型糖度和硬度的预测准确率分别提高了5.87、6.48个百分点,苹果品质回归模型糖度和硬度的决定系数分别提高了0.139 7和0.088,与基于点光源的预训练模型相比达到了更好的效果。更多还原

【Abstract】 Aiming at the problem of large detection errors caused by insufficient sample quantity in traditional fruit and vegetable quality detection methods,an apple quality detection method based on photon transmission simulation under surface light source was proposed. Taking apples as the research object,Monte Carlo method was used to simulate the motion trajectory of photons on the apple doublelayer flat model,totally 20 000 apple tissue surface brightness distribution maps were quickly obtained,optical parameters were used as labels,and input convolutional neural network training was used to obtain the model. The fine-tuning migration was applied to a small number of data sets of measured apple spectral images to realize the inversion of optical characteristic parameters. Finally,the output result of the fully connected layer of the network model was associated with the quality of the apple,so as to realize the non-destructive testing of the sugar content and hardness of the apple. The final result was that the inversion accuracy of pulp absorption coefficient μ_a2 was 93. 24%,and the inversion accuracy of pulp scattering coefficient μ_s2 was 92. 54%. The prediction accuracy of sugar content and hardness of the quality classification model were improved by 5. 87 and 6. 48 percentage points compared with that of the traditional method. The determination coefficient of sugar content and hardness of the quality regression model was improved by 0. 139 7 and 0. 088 compared with that of the traditional method. Compared with the pre-trained model based on point light source,it also achieved better results.更多还原