基于近红外相机成像的红提串缺陷检测

Defect detection method of red globe grapes bunches based on near infrared camera imaging

目的:探寻快速、准确的葡萄串褐斑及损伤腐烂检测方法。方法:通过近红外工业相机采集红提串的彩色图像(RGB)和近红外图像(NIR)。利用近红外图像(NIR)运用Sobel算子提取清晰葡萄串的边缘和褐斑及损伤腐烂部分的边缘,然后将图像通过自适应阈值对图像进行二值化,实现图像分割,通过归一化超绿法和寻找大连通域去除葡萄串边缘及果梗,分别提取红提串缺陷和果粒边缘的圆形度、矩形度和外接矩形长宽比的形状特征参数,并建立基于BP神经网络和支持向量机的缺陷和果粒边缘判别分类模型,通过建立的分类模型实现果粒边缘的剔除,获得褐斑及损伤腐烂的图像信息。结果:利用上述检测方法对60个葡萄串样本进行验证,完好葡萄串判别准确率为90.00%,褐斑及损伤腐烂葡萄串判别准确率为93.33%,综合判别准确率达到91.67%。结论:研究建立的褐斑及损伤腐烂图像的检测方法可以实现红提葡萄的分级挑选。

Objective:The study aimed to explore a fast and accurate method to detect brown spot and damage decay in grape bunches.Methods:Colour images(RGB)and near-infrared images(NIR)of red globe grapes bunches were captured by a near-infrared industrial camera.The edges of the samples and the edges of the defective parts were first extracted by applying the Sobel algorithm to the NIR images(NIR),and then the images were binarized by the adaptive thresholding algorithm to achieve the segmentation of the images.Then the sample edges and fruit stalks were removed by the normalized supergreen method and the finding large connected domain algorithm to extract the shape feature parameters such as roundness,rectangularity and external rectangular aspect ratio of the defective part of red globe grapes bunches and fruit edges,respectively.Finally,a classification model based on BP neural network and support vector machine was developed to discriminate the defective parts and fruit edges.The model enables the rejection of kernel edges to obtain image information of brown spots and damage decay.Results:Using the above-mentioned testing method to verify 60samples,the accuracy of discriminating red globe grape bunches with intact appearance was as high as 90.00%,those with defects reached 93.33%,and the overall discriminating accuracy reached91.67%.Conclusion:The study established a method to detect brown spot and damage decay images to enable grading and selection of red globe grapes.