基于计算机视觉的核桃外观缺陷检测

Walnut Appearance Defect Detection Based on Computer Vision

为快速准确识别核桃外观缺陷(黑斑、破裂),自行搭建图像采集系统采集样本图像。预处理后采用形态学和逻辑运算去除背景,基于样本图像提取18个颜色特征参数和20个纹理特征参数。采用形态学和逻辑运算提取缺陷部分和样本投影像素面积的比值t以及样本图像阈值分割后二值图像的欧拉数。分别采用回归系数法(Regression Coefficient,RC)和连续投影法(Successive Projections Algorithm,SPA)优选特征参数并建立偏最小二乘法(PLS)模型。结果表明,基于SPA法优选特征参数建立的模型性能最优。将SPA法提取的5个优选特征参数作为输入建立最小二乘支持向量机(LS-SVM)模型,并对预测集样本进行预测。结果表明,对正常核桃、黑斑核桃、破裂核桃的判别准确率分别为88.9%、83.3%、94.6%,总判别率为88.9%。本研究建立的方法能够很好的对核桃外观缺陷进行检测,为今后核桃的在线检测分选提供了技术支持。

For rapid and accurate identification of the appearance defects(black spots, ruptures) of walnuts, an image acquisition device was established to collect sample images. After pretreatment, the background was removed by morphological and logical operations. Then, 18 color feature parameters and 20 texture feature parameters were extracted based on the sample images. Morphological and logical operations were used to extract the ratio of the defect portion to the projected pixel area of the sample, as well as the Euler number of the binary image after the sample image threshold was segmented. Regression coefficient(RC) and successive projections algorithm were used to optimize the feature parameters and establish a partial least squares(PLS) model. The results showed that the model established based on the preferred SPA method exhibited the optimal performance. The five preferred feature parameters extracted by SPA method were used as input to establish the least squares support vector machine(LS-SVM) model, which was used to predict the prediction set samples. The results showed that the discrimination of normal walnut, walnut with black spot(s) and cracked walnut was accurate with rates as 88.9%, 83.3% and 94.6% respectively, and the overall accuracy rate of discrimination was 88.9%. The method established in this study can well detect the appearance defects of walnut and provides technical support for online detection and sorting of walnut in the future.