为了确定可用于苹果早期轻微损伤检测的有效波长,以具有代表性的阿克苏苹果为研究对象,采用高光谱成像技术和分段主成分分析方法对损伤发生仅为半小时之内的苹果进行损伤检测研究,对比分析不同光谱区域主成分分析对识别结果的影响,优选出识别光谱区域（780~1000 nm）.基于此光谱区域结合主成分图像权重系数获取2个有效波长（820和970 nm）,并利用这2个波长和全局阈值理论开发了多光谱轻微损伤提取算法.利用独立测试集中25个正常苹果和25个损伤苹果对算法的性能进行评估,结果表明,正常果的识别率为100%,损伤果的识别率为96%,整体检测精度为98%.该研究所获得的有效波长可为开发基于多光谱成像技术的苹果损伤检测系统提供参考.
Bruise is one of the main defects of apple, which could be caused by impact or mechanical damage during harvest and handling stages. Early detection of slight bruises on apples is important for an automatic apple sorting system. A hyperspectral imaging system with the wavelength range of 450-1 000 nm was built for detecting bruises happened in half an hour on ＇Akesu＇apples. The hyperspectral imaging system was used as a powerful tool to determine the effective wavelengths that could be used for the detection of bruises on apples. Principal component analysis （PCA） is a very effective method for data dimension reduction and feature extraction of the hyperspectral data cube. However, too many wavelengths from entire spectrum data were usually used to perform the PCA operation. Therefore, the performance of PCA was degraded due to a lot of noises. In addition, too many effective wavelengths were also not effective to develop a multispectral system. In this study, segmented PCA was used to select the effective wavelengths. First, PCA was conducted on the three spectral ranges 450-780 nm, 450-1 000 nm and 780-1 000 nm. Then, the optimal wavelength region 780-1 000 nm for bruise detection was selected by visually contrasting and analyzing the obtained principal component （PC） images of the PCA on the three different wavelength regions. Two effective wavelengths 820 and 970 nm with weighing coefficients at peaks and valleys were determined using the loading coefficients of the PC2 image of PCA on 780-1 000 nm. The PC2 image obtained from PCA on two effective wavelengths 820 and 970 nm was used for bruise detection. First, the PC2 images were processed by applying the Gaussian blur filter. Then, a bruise detection algorithm based on the two effective wavelengths and a global threshold method was developed. Independent validation set of 25 intact and 25 bruised apples was used to evaluate the performance of the developed algorithm. Results show that 100% of the intact apples are correctly classified, 96% of th
Transactions of the Chinese Society of Agricultural Engineering
defects, image processing, principal component analysis, hyperspectral imaging, effective wavelength,apple