CAJ | 학술논문

利用高光谱成像技术，明确苹果虫害无损检测的最优特征向量，以实现对苹果虫害的快速、准确、无损检测。本文首先对646 nm波长的特征图像进行阈值分割、膨胀与腐蚀运算获得掩膜图像，并利用掩模图像对高光谱图像进行掩模和主成分分析，对获得的PC1（the first principal component，第一主成分）图像进行最大熵阈值分割以有效提取虫害区域。然后对比分析虫害区域与正常区域图像的纹理特征，提取灰度共生矩阵的4个方向的4个纹理参数（能量、熵、惯性矩和相关性），并且采用基于高光谱图像的光谱差值获取了2个特征波长对应的光谱相对反射率作为光谱特征。优化组合纹理特征和光谱特征成4个特征向量组，采用BP（back propagation，反向传播）神经网络对正常苹果和虫害苹果进行检测。结果表明，融合0度方向的能量、熵、惯性矩和相关性的纹理特征和646、824 nm波段的相对光谱反射率的光谱特征进行检测识别效果最好，正常果的识别率为100%，虫害果的识别率为100%，并且速度快、误差小。该研究所获得的特征向量可为开发多光谱成像的苹果品质检测和分级系统提供参考。
이용고광보성상기술，명학평과충해무손검측적최우특정향량，이실현대평과충해적쾌속、준학、무손검측。본문수선대646 nm파장적특정도상진행역치분할、팽창여부식운산획득엄막도상，병이용엄모도상대고광보도상진행엄모화주성분분석，대획득적PC1（the first principal component，제일주성분）도상진행최대적역치분할이유효제취충해구역。연후대비분석충해구역여정상구역도상적문리특정，제취회도공생구진적4개방향적4개문리삼수（능량、적、관성구화상관성），병차채용기우고광보도상적광보차치획취료2개특정파장대응적광보상대반사솔작위광보특정。우화조합문리특정화광보특정성4개특정향량조，채용BP（back propagation，반향전파）신경망락대정상평과화충해평과진행검측。결과표명，융합0도방향적능량、적、관성구화상관성적문리특정화646、824 nm파단적상대광보반사솔적광보특정진행검측식별효과최호，정상과적식별솔위100%，충해과적식별솔위100%，병차속도쾌、오차소。해연구소획득적특정향량가위개발다광보성상적평과품질검측화분급계통제공삼고。
Insect pestilence is one of the main defects of the apple industry, which could be caused by pest entrance during apple tree growth stages. Insect pest detection in apples is important for an automatic apple quality inspection and sorting system. In this study, we intended to determine the feature vectors that can be used for nondestructive detection of apple fruit insect pests and utilized hyperspectral imaging technology to carry out an effective method for rapid, non-invasive detection of the intact apples and insect pests. There were 160 samples of 80 intact and 80 insect infected ‘red Fuji’ apples to be investigated from an apple planting demonstration garden in the Shenbei New Area in Shenyang city. A hyperspectral imaging collection system with the wavelength range of 400-1 000 nm was established to acquire the hyperspectral images of these apple samples. Via the analysis of spectral reflectance of apple pest parts and the normal region, there were obvious differences in spectral reflectance at the 646 nm wavelength. So, the image of the 646nm wavelength was named the feature image. Then, the feature image was manipulated by threshold segmentation, dilation, and erosion operation, to obtain a mask image. The mask image was used for image analysis to mask and carried on principal component analysis. The optimum PC1 image was chosen and handled by the maximum entropy threshold segmentation to extract the pest region. Later, a comparative analysis of the texture feature of the insect infested region and the normal region on apples of the PC1 image, a region of interest (ROI) with 80 pixels×65 pixels of the PC1 image of each sample, was obtained. The texture features of the gray level co-occurrence matrix (of energy, entropy, moment of inertia and correlation) in four directions, which were 0, 45, 90, and 135 deg, respectively, were extracted. In addition, the spectral relative reflectance of the apple surface pests and normal regions, whether it was visible or near infrared region, had obvious difference. So the two spectral features of the spectrum relative reflectivity at 646 and 824 nm wavelength were acquired, which had larger relative reflectance differences between the apple surface pests and normal regions in the visible region and near infrared region, respectively. Feature vector selection was one of the key steps in detecting apple insect pests. For faster and more accurate detection of the apple insect pests, in this study, the optimization and integration of the texture features and the spectral feature vectors was analyzed. Four feature vector groups were posed respectively as the input vector of the BP neural network. The validation set of 30 normal apples and 30 insect infested apples was detected by using the BP neural network. The recognition rate was the highest when there was a fusion of the texture features of energy, entropy, moment of inertia, the correlation of 0 deg direction, and the spectral features of relative spectral reflectance with two feature wavelengths of 646 and 824 nm. A recognition rate of the normal apples and insect infested apples was 100 percent. Besides, in this case, the speed of detection is the fastest, and the MSE error is the smallest. Results show that the obtained feature vectors based on hyperspectral imaging technology can identify insect infestation effectively and provide a reference for apples quality detection and grading system using multispectral imaging.