基于图像特征识别的马铃薯薯皮粗糙度分级研究

Potato Tuber Skin Roughness Classification Analysis Based on Image Characteristics Recognition

【目的】马铃薯薯皮粗糙度分级研究可以提供块茎外观品质性状无损检测方法,为客观评价品质质量和高通量筛选品种提供理论和实践基础。【方法】以79份马铃薯品种(系)为供试材料,利用相机采集有/无芽眼的薯皮图像。基于MATLAB R2016a软件对薯皮图像预处理,随机选择8份材料用相关函数指标比较图像灰度化、增强及去噪效果。利用灰度共生矩阵(gray level co-occurrence matrix,GLCM)提取图像特征参数角二阶矩(angular second moment,ASM)、熵(entropy,ENT)、对比度(contrast,CON)和相关度(correlation,COR),并确定矩阵最适像素距离(d)。比较两类薯皮图像特征参数间的差异,选择差异较小的薯皮图像特征集进行统计分析和分类识别。构建支持向量机(support vector machines,SVM)和BP神经网络(backpropagation neural network,BPNN)模型对薯皮粗糙度分级分类,模型分级精度评价指标为准确率、精准率、召回率及调和平均数。【结果】加权平均值法进行灰度处理后的薯皮图像纹理结构清晰,图像清晰度评价值为2.5698±0.5959,显著高于平均值法(1.8035±0.4856)和最大值法(1.0535±0.4088);直方图均衡化增强后的薯皮图像灰度级范围由100—200扩大为0—200,灰度分布更加广泛;中值滤波对3×3窗口下的薯皮图像椒盐噪声去噪效果明显,峰值信噪比(peak signal-to-noise ratio,PSNR)最大((28.6250±3.9784)Bp),显著高于3×3和5×5窗口下对高斯噪声去噪后的PSNR。通过GLCM(d=4)提取的两类薯皮图像特征参数间差异显著,选择其中差异较小的无芽眼薯皮图像特征集进行统计分析和分类识别,结果表明该特征集变异系数差异明显,对比度变异系数最大(0.40),其次是角二阶矩(0.24)和相关度(0.23),熵变异系数最小(0.18)。将该特征集作为分类模型输入变量用于薯皮分类,相较于BP神经网络,SVM对马铃薯薯皮粗糙度的整体分类性能较高,准确率为87.5%。其中,对光滑皮和重麻皮的预测准确度和识别能力最高,精准率均为100%,召回率分别为85.7%和100%,调和平均数分别为92.3%和100%。【结论】综合利用本研究提出的图像处理技术及GLCM提取的纹理特征参数能有效表征马铃薯块茎薯皮粗糙度差异;通过构建SVM分类模型可实现基于机器视觉的马铃薯薯皮粗糙度分级,且准确率达87.5%。

【Objective】The classification analysis of potato tuber skin roughness could provide the non-destructive testing methods for tuber appearance quality traits,which would establish the theoretical and practical base for the objective evaluation of tuber quality and high-throughput screening varieties.【Method】Seventy-nine potato varieties(lines)were selected as materials,and the images of tuber skin with and without bud-eyes were taken by camera.The tuber skin images were preprocessed using MATLAB R2016a software.Eight materials were randomly selected to compare the effect of image graying,enhancement and denoising using the correlation function indicators.The image characteristic parameters,angular second moment(ASM),entropy(ENT),contrast(CON)and correlation(COR)were extracted using the gray level co-occurrence matrix(GLCM),and the suitable distance(d)of GLCM were determined.The differences in two types of tuber skin image feature parameters were compared,and the set of tuber skin image features with less difference was selected for statistical analysis and classification recognition.The support vector machine(SVM)and backpropagation neural network(BPNN)models were constructed for tuber skin roughness classification,and the evaluation indexes of model grading accuracy were accuracy,precision,recall and harmonic mean,respectively.【Result】The texture structure of tuber skin image after grayscale processing using the weighted average method was clear,and the evaluation value of image clarity was 2.5698±0.5959,which was significantly higher than that of the mean method(1.8035±0.4856)and the maximum method(1.0535±0.4088).The gray scale range of tuber skin image after histogram equalization enhancement was expanded from 100-200 to 0-200,which made the gray distribution wider.The salt noise denoising effect of tuber skin images using the median filter under 3×3 sliding windows was obvious,and the peak signal-to-noise ratio(PSNR)was maximum(28.6250±3.9784 Bp),which was significantly higher than that under 3×3 and 5×5 windows.Two types of tuber skin image feature parameters extracted by GLCM(d=4)were significantly different,and the set of tuber skin image features(without bud-eyes)with less difference was selected for statistical analysis and classification recognition.The results indicated that the variation coefficient of these parameters was varied significantly.The variation coefficient of contrast was the largest(0.40),followed by the angular second moment(0.24)and correlation(0.23),and the variation coefficient of entropy was the smallest(0.18).Using the feature set as the input variable of tuber skin classification model,the overall classification performance of SVM was higher than BP neural network,and the accuracy reached 87.5%.Especially,the prediction accuracy and recognizability of SVM for smooth and heavy hemp skins was the highest.The accuracy reached 100%,the recall reached 85.7%and 100%,and the harmonic mean reached 100%and 92.3%,respectively.【Conclusion】The combination of the image processing techniques presented in this study and the GLCM extracted texture feature parameters could effectively characterize potato tuber skin roughness variations.The tuber skin roughness grading based on machine vision could be achieved by constructing SVM classification model,and the accuracy reached 87.5%.