Whether from the perspective of agricultural production or food safety, potato germination detection is of great significance. Since the features (color, texture and context) of the germination area are similar to those of the non-germination area, the existing vision frameworks are difficult to accurately detect the germinations on the surface of potatoes. In this study, the method for detecting potato germination based on multispectral image combined with supervised multiple threshold segmentation model (SMTSM) and Canny edge detector was proposed. The SMTSM based on Genetic Programming algorithm combined with a hybrid fitness function (HF-GP) was utilized to transform the original multispectral images into multiple 2-D images for improving the contrast between region of interest (ROI) and background. A sub-mask of each transformed image was constructed using optimal segmentation threshold, and all of sub-masks were merged through pixel-multiplication to obtain segmentation mask. Meanwhile, in order to filter out the boundless areas that are misidentified as germinations, Canny edge detector was used on gray image to obtain edge mask. Finally, the segmentation mask and the edge mask were combined to complete the detection of germination of potato. Experimental results shown that the proposed method achieved the TPR of 90.91% and the precision of 89.28% for the edible potatoes, which were 4.17–19.05% and 12.39–24.62% higher than the competitive detectors in TPR and precision respectively. For the breeding potatoes, the proposed method with 89.67% of TPR and 86.37% of precision was 9.74–24.58% and 15.70–20.39% better than the competitors in TPR and precision respectively. The comparison confirms the proposed method has excellent detection effect on potato’s germination.

无论从农业生产还是食品安全的角度来看，马铃薯发芽检测都具有重要意义。由于发芽区域的特征（颜色，纹理和背景）与非发芽区域的特征相似，因此现有的视觉框架很难准确地检测马铃薯表面的发芽。提出了基于多光谱图像结合监督多阈值分割模型（SMTSM）和Canny边缘检测器的马铃薯萌发检测方法。利用基于遗传规划算法结合混合适应度函数（HF-GP）的SMTSM将原始多光谱图像转换为多个二维图像，以改善感兴趣区域（ROI）和背景之间的对比度。使用最优分割阈值构造每个变换图像的子掩模，并且通过像素相乘合并所有子掩模以获得分割掩模。同时，为了滤除被误认为是发芽的无边区域，在灰色图像上使用了Canny边缘检测器以获得边缘蒙版。最后，将分割蒙版和边缘蒙版结合起来，完成对马铃薯萌发的检测。实验结果表明，所提方法达到了 将分割蒙版和边缘蒙版结合起来，完成对马铃薯萌发的检测。实验结果表明，所提方法达到了 将分割蒙版和边缘蒙版结合起来，完成对马铃薯萌发的检测。实验结果表明，所提方法达到了食用马铃薯的TPR为90.91％，精度为89.28％，分别比同类竞争产品的TPR和精度高4.17–19.05％和12.39–24.62％。对于养殖土豆，所提出的方法与89.67％TPR和86.37％，精确度为9.74-24.58％，比竞争对手15.70-20.39％更好TPR和精度分别。比较结果表明，所提出的方法对马铃薯发芽具有很好的检测效果。