Segmentation of Crop Disease Images with an Improved K-means Clustering Algorithm

Abstract

Abstract. Disease spot segmentation from crop leaf images is a key prerequisite for disease early warning and diagnosis. To improve the accuracy and stability of disease spot segmentation, an adaptive segmentation method for crop disease images based on K-means clustering is proposed. The approach is based on three stages. First, the excess green feature and the a* component of the CIE (L*a*b*) color space were combined to adaptively learn the initial cluster centers. Second, iterative color clustering of two clusters was conducted using the squared Euclidian distance as the similarity distance. Finally, the distance of a* components between two clusters as the clustering criterion function was used to correct the clustering results. To verify the effectiveness of the proposed method, segmentation experiments were performed on images of three kinds of cucumber diseases and one kind of soybean disease. The results of the experiments were compared with the results obtained using a fixed threshold method, the Otsu method, the traditional K-means clustering method, and the Renyi entropy method, which showed that our adaptive segmentation method was accurate and robust for segmentation of crop disease images. Keywords: Adaptive, CIE L*a*b*, Disease spot, Image segmentation, K-means clustering.