Pixel based bruise region extraction of apple using Vis-NIR hyperspectral imaging
Pixel based bruise region extraction of apple using Vis-NIR hyperspectral imaging
复制标题
DOI：
10.1016/j.compag.2018.01.013
复制
发表时间：
2018-03-01
影响因子：
8.3
通讯作者：
Liu, Yangyang
中科院分区：
文献类型：
--
作者：
Che, Wenkai;Sun, Laijun;Liu, Yangyang
苹果上的瘀伤会导致苹果内部腐烂和外观瑕疵，直接影响苹果的保鲜和销售。因此，本文提出了一种有效的基于像素的擦伤区域提取方法，以获取完整的擦伤区域。利用高光谱成像(HSI)系统分别在损伤实验后0、12和18h获取了60个苹果的高光谱图像。利用主成分分析(PCA)对高光谱图像立方体的数据量进行压缩，剔除冗余数据。在按一定规则选择感兴趣区域(Rod)后，建立了不同的基于像素的苹果瘀伤提取模型，并进行了比较。结果表明，随机森林模型具有较高且稳定的分类精度，说明随机森林算法比其他算法更适合于苹果擦伤的分类。伤痕提取模型的平均准确率达到99.9%。与目前文献中提取苹果擦伤最常用的图像处理方法相比，RF模型预测的擦伤区域与真实的擦伤区域更加一致。此外，通过分析所建立的射频模型的特征重要性分数，挑选出与擦伤区域相关的675 nm和960 nm附近的两个特征波段来降低数据的维度。结果表明，基于高光谱图像的苹果完整损伤区域检测方法对于提高苹果分级分选的效率具有很大的潜力。
Bruises on apples will directly influence its preservation and marketing for they can cause the internal decomposition and flaws of the appearance of apples. Therefore, an effective pixel based bruise region extraction method was proposed in this study to obtain the complete bruise region. Hyperspectral images of 60 apples were obtained via the hyperspectral imaging (HSI) system at 0, 12 and 18 h after the damage experiment. Principal Component Analysis (PCA) was used to compression data size and eliminating redundant data of hyperspectral image cubes. After the selection of the region of interest (ROD by certain rules, different pixel based apple bruise extraction models were built and compared. The result shows that Random Forest (RF) model have a high and stable classification accuracy, which turns out that RF algorithm is more suitable for classifying bruises on apples than others. The average accuracy of bruise extraction models reached 99.9%. Compared with the most used image processing method in recent literature for extracting bruises of apples, the bruising region predicted by RF model was more consistent with the true bruise region. Additionally, two characteristic wavebands around 675 nm and 960 nm related to the bruise region were singled out for reducing the dimensionality of data by analyzing the feature importance scores of the built RF model. The overall results indicated that the proposed method has a great potential to detect complete bruise region on apples based on hyperspectral imaging for improving the efficiency of apple grading and sorting.