基于修正聚类分析与核密度估计的彩色图像分割

Color image segmentation based on clustering adjustment and kernel density estimation

由于欧氏距离在度量向量之间的相似性时仅具有平移不变性,直接运用像素向量欧氏距离的聚类划分结果进行高斯核密度估计实现彩色图像分割时,存在相同区域系数分布不均匀、边界不清晰等问题,而具有旋转不变性和尺度不变性的余弦距离能修正欧氏距离的不足。所以,在像素向量欧氏距离聚类划分之后辅以余弦相似性修正,以提高聚类划分的鲁棒性;再对修正的聚类划分结果进行改进的快速高斯变换(IFGT)实现核密度估计,以提高彩色图像分割的抗噪性能。针对噪声分布复杂的打印扫描彩色图像进行分割实验,结果表明,该方法提高了抗噪声性能,改善了相同区域系数分布的均匀性。

Because the cosine similarity is invariant to rotation and scale, which is the Euclidean distance deficient of, the adjustment of the dusters obtained by the k-center farthest clustering method was proposed based on the cosine similarity of color pixel vectors. To adaptively select the truncation number of Improved Fast Gaussian Transform （IFGT）, the adjustment result was used to the IFGT algorithm to scgment the color image. For the validation of the proposed method, a printed and scanned color image was segmented. The result shows that the uniformity of coefficients distribution in the same region is improved.