基于激光散斑的梨缺陷与果梗/花萼的识别

为了测试利用激光散斑技术区分梨的缺陷与果梗/花萼的可行性,建立了激光散斑图像采集系统,对皇冠梨缺陷(腐烂)部位以及完好部位(花萼/果梗,无缺陷部位)分别进行了激光散斑图像的采集。利用Fujii方法(Fujii’s method)和加权广义差分方法(weighted generalized differences,WGD)对512幅散斑图像进行分析,对得到的Fujii和WGD结果图进行灰度共生矩阵特征提取,分别提取了角二阶矩、熵、惯性矩和相关性相应的均值及标准差,共计16组特征量。利用ROC曲线(receiver operator characteristic curve,ROC)进行特征量选取,结合约登指数测试单一特征量的分类效果,并利用二元logistic回归方法对所选特征量两两组合进行分析,结果显示基于WGD方法得到的角二阶矩均值与相关性标准差相结合在区分缺陷时效果最好,建模和预测准确率均达到了97.5%。试验的结果表明利用激光散斑图像方法对梨缺陷与果梗/花萼进行识别是可行的。

Pear defect and stem/calyx discrimination using laser speckle

Abstract: Laser speckle is a new technology for non-destructive detection in agriculture. When laser light with good coherence irradiates on the optically rough surface or the surface with some kind of activity, the scattering light will interfere with each other and form a mottled pattern called laser speckle. The laser speckle contains various information of the reflector, such as the roughness, the particle's motion and the temperature information. Huangguan pears were used as the object in this research to investigate the possibility of identifying the defects from the stem/calyx and the sound area of pears based on laser speckle technology. A laser speckle imaging system was established which contained a semiconductor laser (635 nm, 50 mw) applied as the light source and a digital signal generator applied as the trigger source of the CCD (charge coupled device) camera. Two hundred pears including one hundred sound pears and one hundred pears with the defect (rot) were tested. The speckle images of defect (rot) parts and good parts (calyx/stem, sound area) of the pears were collected under the same condition. First of all, the speckle images were converted into the grayscale images using Matlab 2011b. The method of Fujii and the weighted generalized differences method (WGD) were used to analyze the grayscale speckle images to get the images of Fujii and WGD. Then gray level co-occurrence matrix (GLCM) was used to extract the mean and the standard deviation values of the angular second moment (ASM), entropy (ENT), moment of inertia (INE) and correlation (COR) from the images of Fujii and WGD, respectively. Therefore, in total, 16 features were extracted. The performance of each feature was evaluated by the receiver operator characteristic (ROC) curve. According to the ROC curves, the features whose values of the area under the curves (AUC) were higher than 0.5 were chosen for further analysis. The best threshold value of each selected feature was calculated by Youden's index. The classification analysis based on single feature was tested using the best threshold value. Besides, the classification analysis based on multiple features was carried out by the binary logistic regression. The combination of every two features was tested to get better classification accuracy. The results showed that there were seven features whose AUC values were bigger than 0.5. In classification of single feature, the ASM extracted from WGD image had the best performance whose overall accuracies of calibration and validation sets were 96.4% and 96.7% respectively. In classification of multiple features, the combination of the ASM and standard deviation of correlation extracted from WGD image had the best performance whose overall accuracies of calibration and validation sets were both 97.5%. In order to find out the causes of the error, the original RGB (red, green and blue) images of the misjudged samples were studied. It turned out that the defect area which was misjudged to be normal was not obvious. Besides, the defect area was larger than the light spot whose diameter was 20 mm which could not cover the defect area completely. Therefore, the texture features of laser speckle images of misjudged samples were closer to the sound area than the defect area, which led to the miscalculation. This research shows that the method of laser speckle imaging is feasible in the detection of pear defect (rot) and stem/calyx.