基于K-means聚类的柑橘红蜘蛛图像目标识别

Citrus red mite image target identification based on K-means clustering

为快速检测红蜘蛛虫害,该研究采用基于Lab颜色模型中a(红/绿)、b(黄/蓝)层信息的K-means聚类法识别彩色图像中的红蜘蛛。试验选取8幅具有不同清晰度的柑橘红蜘蛛图像,采用基于Sobel边缘检测算子的评价函数计算图像清晰度评价值以评价图像清晰度,对比采用灰度法和包含2、3、4或5个聚类中心的K-means聚类法的目标识别效果和识别效率。结果表明,灰度法对8幅图像中红蜘蛛目标识别率平均值为29%,误判率平均值为201%,无法应用于复杂背景图像中的红蜘蛛目标识别。包含5个聚类中心的K-means聚类法对清晰度较高的图像识别率为100%,误判率为0,对清晰度较低的图像识别率为88%,误判率为0;当图像尺寸较小时,包含4个聚类中心的K-means聚类法识别效率与灰度法相当;当图像尺寸较大时,重复计算聚类中心导致识别耗时较长;基于Lab颜色空间的识别算法无法有效识别其他颜色的红蜘蛛,继续研究的方向为引入红蜘蛛形态信息以提高识别准确率和优化聚类中心的选取以降低识别耗时。

In recent years, the harm of citrus red mites in orchard is gradually becoming serious. Research on correctly and effectively identifying the occurrence of citrus red mites in orchard, thus spraying pesticide reasonably, will have important significance on the achievement of orchard harvest and environment protection. This study aims at providing an effective method for rapid citrus red mites pest detection from color images with complex background. The proposed algorithm took advantage of an Lab modeled image＇s a （red and green） and b （yellow and blue） component rather than RGB modeled image used by traditional image processing algorithms. The citrus red mite targets within an Lab modeled image were indentified based on K-means clustering method. Experiments were conducted using 8 citrus red mite images with different image clarity to compare the target recognition effect and efficiency of the grayscale method and K-means clustering with 2, 3, 4 and 5 cluster centers. The evaluation indicators for the comparison were recognition accuracy, error rate and identification time consumption. The recognition accuracy is the ratio of the number of correctly recognized red mites by algorithm to the number of actual red mites. The error rate is the ratio of the number of falsely recognized red mites by algorithm to the number of actual red mites. The image clarity was evaluated using the image clarity evaluation index which was calculated by the evaluation function based on Sobel edge detection operator. Results indicated that： although short time consumption was achieved, it was invalid to use the grayscale method for red mite recognition under complex background images, since the averaged recognition accuracy and error rate obtained from using this method on the 8 images was 29% and 201%, respectively. The recognition accuracy was 100% and the error rate was reduced to zero using K-means clustering with 5 cluster centers for images with higher clarity while an 88% recognition accuracy and zero error rate was achieved for images with lower clarity. Time consumed by K-means clustering with 4 cluster centers was similar to the time consumed by the grayscale method. Along with a grown image size, repeated cluster center calculation led to longer time consumption.It was hard to identify red mite with other color rather than that with red within the Lab color space, thus further research should focus on improving recognition accuracy by introducing red mites shape information. Furthermore, to decrease time consumption by optimizing cluster center selection will be another research direction.