【摘要】 图像数据中同一目标区域相邻像素点之间具有强相关性、存在噪声、数据规模大等特征，期望分割的结果在图像空间上连续。针对上述提出的图像数据的特征，提出了一种基于层次聚类的图像分割算法。首先，采用简单线性迭代聚类（SLIC）超像素分割算法对图像预处理，把图像分割为若干个超像素，提取超像素的特征，基于超像素间的邻接关系构造区域邻接图（RAG）；然后，采用7种不同的类间差异计算方法，引入邻接区域合并、最小类间差异合并机制及最小类内差异（MID）准则对超像素进行层次聚类。在Pascal VOC数据集上进行实验，实验结果表明：与SLIC相比，该算法其边界召回率（BR）提高了0.180，而欠分割误差（UE）降低了0.088。更多还原

【Abstract】 Image data has the characteristics of strong correlation between adjacent pixels in the same target area, noise, large data scale, etc., and the result of segmentation is expected to be continuous in the image space. Aiming at the characteristics of the image data proposed above, an image segmentation algorithm based on hierarchical clustering was proposed. Firstly, the Simple Linear Iterative Clustering(SLIC) superpixel segmentation algorithm was used to preprocess the image. The image was divided into a number of superpixels. And the features of the superpixels were extracted. Region Adjacency Graph(RAG) based on the adjacency relationship between the superpixels was constructed. Secondly, seven different calculation criteria for inter-class difference were adopted. And the adjacent region merging, the minimum interclass difference merging mechanism and the Minimum Internal Difference(MID) criterion were introduced to perform hierarchical clustering of superpixels. Experiments on the Pascal VOC dataset show that the proposed algorithm has the Boundary Recall(BR) 0. 180 higher and Under-segmentation Error(UE) 0. 088 lower than the SLIC.更多还原