Abstract: It is of great significance to combine large-area homogeneous regions with superpixels to detect, track and recognize targets and to process remote sensing images. In the process of merger, superpixel is required to have good edge retention. The traditional method of superpixel segmentation pursues shape rules within ignoring edge fit. In view of this, an image segmentation method based on RGB three-dimensional histogram combined with DBSCAN is proposed. Firstly, 3D RGB histogram of image is analyzed to obtain initial superpixels with high edge conformance, then the proper eigenvalues are selected, and the superpixels are merged by using the DBSCAN algorithm to generate a large homogeneous region. Experimental results show that the edge preserving and computing efficiency of new method are better than those of traditional methods. When using DBSCAN to combine superpixels, the segmentation accuracy is obviously improved and homogeneity area edge is more exact.

Key words: RGB 3D histogram, superpixels, homogeneous regions, DBSCAN clustering

摘要： 合并超像素生成大面积同质区对目标检测、跟踪和识别及遥感影像处理具有现实意义。在合并过程中，要求超像素具有良好的边缘保持性，传统的超像素分割方法追求形状规则而忽略边缘的贴合度。有鉴于此，提出一种基于RGB三维直方图结合DBSCAN的图像分割方法。首先分析图像三维RGB直方图获取边缘贴合度很高的初始超像素，进而选择适当的特征值利用DBSCAN算法对超像素合并以生成较大同质区。实验证明：新方法获取超像素的边缘保持性和运算效率都优于传统方法，采用DBSCAN合并超像素时，其分割精度有明显提升，而且同质区边缘更加准确。

关键词: RGB三维直方图, 超像素, 同质区, DBSCAN聚类