【摘要】 目前高光谱成像技术已成为蓝莓品质自动检测的重要手段，其中提取特征波长是一个重要步骤。为获得最佳的腐烂蓝莓高光谱特征波长图像，提出一种基于深度学习的高光谱图像特征波长图像选取方法。首先提取腐烂蓝莓高光谱各波长图像的高频分量图像，然后提出利用深度学习图像质量评价网络（DIQA）选取腐烂蓝莓高光谱图像中的最佳图像。在DIQA第一阶段对腐烂蓝莓高光谱图像的高频图学习误差图。在DIQA第二阶段将第一阶段能够预测误差图的网络作为对腐烂蓝莓高光谱图像质量评价的骨干网络，并对网络做全局平均池化，最后引入两个特征以弥补信息量损失，通过评分评价图像质量。结果表明：通过对400～1000nm的472个波长下蓝莓高光谱图像进行质量评价，得到最佳图像质量的5个波长(664,721, 836,854,884nm)。与PCA提取的特征波长(454, 607, 699,913,967nm)图像相比较得知，所提取的腐烂蓝莓特征波长图像具有较多优势。另外采用卷积神经网络ResNet50对DIQA与PCA构建的蓝莓特征波长图像数据集进行学习，得出DIQA选取的腐烂蓝莓特征波长图像构建的数据集，损失能较快的收敛，并且在验证集中能保持稳定的准确率,识别效果最好，最终识别率为99.4%。说明基于DIQA的腐烂蓝莓高光谱特征波长图像选取方法是可行的，可为高光谱图像选取特征波长图像提供一种新的参考方法。更多还原

【Abstract】 At present, hyperspectral imaging technology has become an important means of automatic detection of blueberry fruits quality, in which extracting characteristic wavelength is an important step. In order to obtain the best hyperspectral characteristic wavelength image of decayed blueberry, a method of extracting characteristic wavelength image from decayed blueberry hyperspectral image using deep learning image quality evaluation network(DIQA) was proposed in this paper. Firstly, the high-frequency components of the decayed blueberry hyperspectral images were extracted. Then a deep learning image quality assessment network(DIQA) was proposed to select the best image in the hyperspectral image of decayed blueberries. In the first stage of DIQA, the highfrequency images of decayed blueberries were chosen to learn the error map. In the second stage of DIQA, the network which can predict the error map in the first stage was used as the backbone network to evaluate the quality of decayed blueberry hyperspectral image and global average pooled this network. Finally two features were introduced to compensate for the loss of information, and the image quality was evaluated by scoring. By evaluating the quality of blueberry hyperspectral images at 472 wavelengths from 400nm to1000nm, five wavelengths(664, 721, 836, 854, 884nm) of the best characteristic image quality were obtained. Compared with the characteristic wavelength(454,607,699,913,967nm) image extracted by PCA, the results show that the characteristic wavelength image of decayed blueberries selected by DIQA proposed in this paper had many advantages. In addition, the convolutional neural network ResNet50 was used to learn the blueberry characteristic wavelength image data set constructed by DIQA and PCA. It is concluded that the data set constructed by the decayed blueberry characteristic wavelength image selected by DIQA can converge quickly, can maintain a stable accuracy in the verification set, have the best recognition, and the final recognition rate was 99.4%. The test results show that the proposed selection method of hyperspectral characteristic wavelength image of decayed blueberry based on deep learning DIQA is feasible, which can provide a new reference method for the selection of characteristic wavelength of hyperspectral image.更多还原