基于图像分析的玉米籽粒损伤检测研究

为深入探索玉米籽粒机械损伤的自动、准确、快速识别技术,本试验首先采用灰度法、色度阈值法、色彩恢复的多尺度Retinex法、基于卷积的Sobel法分别对玉米籽粒进行区域分割和质量比较;其次提取玉米籽粒的二值化图像特征,建立玉米籽粒的机械损伤判别分析和逐步剔除模型;最后利用构建和验证集样本对2种模型进行验证。结果表明,基于卷积的Sobel法二值化图像质量最优,其均方误差、峰值信噪比、熵、平均梯度分别为1.813 5、45.545 5 dB、2.838 7 bit/pixel、7.358 4;利用置信区间法得到了正常与机械损伤样本形态特征的最优阈值,各形态指标对判别是否产生机械损伤的贡献程度由大到小依次为面积、周长、最短费特雷直径、最长费特雷直径,其权重系数分别为0.299 5、0.283 2、0.241 7、0.175 5;得到了玉米籽粒多元线性机械损伤判别模型,相关性为0.805,判别分析和逐步剔除模型的平均准确率分别为93.00%、85.67%,构建集与验证集准确率差值分别为2.00%和3.33%。本研究可为玉米籽粒品质视觉检测提供理论依据。

Research on Corn Kernels Damage Detection Based on Image Analysis

To further explore the mechanical damaged identification technology of corn kernels, which is automatic, accurate and rapid, gray scale method, hue threshold method, multi-scale retinex with color restoration and sobel based convolution method were used to segment corn kernels, to access the comparative segmentation quality. The binary image features of corn kernels were extracted to establish the discriminant analysis and progressive elimination recognition model of corn kernels mechanical damaged. The two models were verified by using the construction and validation set of samples. The results showed that the binary image quality of sobel based convolution method was optimal, with the mean square error, peak signal-to-noise ratio, entropy, and average gradient of 1.813 5, 45.545 5 dB, 2.838 7 bit/pixel and 7.358 4, respectively. The confidence interval method was used to obtain the optimal threshold of morphological characteristics of normal and mechanical damaged samples. The contribution degrees of each morphological index to the mechanical damage showed in the order of area, perimeter, shortest Fetrey diameter and longest Fetrey diameter, and their weight coefficients were 0.299 5, 0.283 2, 0.241 7 and 0.175 5, respectively. The multivariate linear mechanical damage discrimination model of corn kernels was obtained, and its correlation was 0.805. The average accuracy of the discriminant analysis model and the progressive elimination recognition model were 93.00% and 85.67%. Accuracy of the construction set and the verification set were 2.00% and 3.33%, respectively. This study can provide a technique basis for the visual inspection of corn grain quality.