基于K均值聚类和开闭交替滤波的黄瓜叶片水滴荧光图像分割

为了监测温室黄瓜叶片湿润情况以计算叶片湿润时间并用于病害预警，利用K-均值聚类算法实现黄瓜叶片的水滴荧光图像分割。选择人工气候室培育的健康且洁净的黄瓜叶片作为试验试材，采用移液枪向叶面、叶缘部位上滴水，模拟不同的叶片湿润情形，使用荧光成像仪蓝光镜头在白天（07：00）和夜晚（18：00）分别采集图像。应用 K-均值聚类算法在L*a*b颜色空间对水滴图像进行分割，首先要将原始图像由RGB颜色空间转换到L*a*b颜色空间，然后在在L*a*b颜色空间中利用a*b*二维数据空间的颜色差异，以欧式距离度量像素间的相似度，使用K均值对图像进行聚类，聚类得到的图像灰度化后进一步用数学形态学中的开闭交替滤波方法进行校正，最终完成图像分割。利用该方法对10幅含有不同水滴数量的黄瓜叶片荧光图像进行分割，为了验证该方法的有效性，分别采用基于H分量直方图分割算法、主动轮廓即C_V模型分割方法、融合K均值聚类和Ncut算法作对比试验。试验结果表明，该方法的平均匹配率、误分率相较于其他3种方法有明显的优势，平均匹配率为81.27%、平均误分率为9.57%，较之于其他3种方法，平均匹配率分别提高了44.11、11.50、10.90百分点，平均误分率分别降低了23.03、5.47和5.05百分点。该方法能够较为准确地将水滴从图像中分割出来，这为用计算机器视觉的方法监测黄瓜叶片的润湿时间提供了新的思路。

Water droplets fluorescence image segmentation of cucumber leaves based on K-means clustering with opening and closing alternately filtering

Abstract: Monitoring moisture condition of cucumber leaves is to calculate leaf wetness duration for the disease forecasting in greenhouse, which is especially important for improving the yield and quality of agricultural products. K-means clustering with opening and closing alternately filtering algorithm was used for the fluorescence images segmentation of water droplets on cucumber leaves. The healthy and clean cucumber leaves in the artificial climate chamber were chosen as the experimental materials. In the experiment, we used a pipette with different volume of water (100 or 200 mL) to drop water on cucumber leaves. Each time, water was dropped to different parts of the cucumber leaves, including leaf surface and margin to simulate different leaf wetness situation. We used the fluorescence imaging instrument to collect the image at day and night. In this article, the image segmentation method was divided into two parts, which included K- means clustering and opening and closing alternately filtering. The main steps of segmentation algorithm of water droplet fluorescence image were as follows. The original images were collected in RGB color space, but the color distribution of the RGB color space was uneven. The advantages of the L*a*b* color space could make up for the shortage. So the original image was firstly converted to the L*a*b* color space from RGB color space. In the L*a*b* color space, all color information was contained in a* and b* components. Secondly the color difference between the two-dimensional data space of a* and b* was used, and Euclidean distance was chosen to measure the similarity between pixels. The fluorescence images were clustered by K- means. After finish of the clustering operation, the images were grayed, and then they were corrected by use of mathematical morphology methods. For morphology methods operation steps, open operation was firstly applied and then close operation was applied. The operations were repeated until the desired results were obtained. And finally the image segmentation was completed. The experiment was carried out to segment ten fluorescence images of cucumber leaves with different numbers of water droplets. In order to verify the validity of the method, we compared our results with three other segmentation algorithms based on H component, active contour model (C_V model), fusion of K-means and Ncut. The results showed that the average matching rate was 81.27% and the average misclassification rate of this method was 9.57%. Compared with the three methods, the average matching rate from our method was improved by 44.11%, 11.50% and 10.90%, respectively. In comparison to the three methods, the average misclassification rate of the method was reduced by 23.03%, 5.47% and 5.05%, respectively. From the experimental data, the segmentation results of the fluorescence images were satisfactory. This method can be used to segment water droplets from the fluorescence images of water droplets on cucumber leaves accurately, which provides a new way to monitor the wetness duration of cucumber leaves by computer vision.