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

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.