基于均值漂移的非线性尺度空间滤波

Non-linear scale space filtering based on mean shift

为解决均值漂移滤波核带宽的选择问题,分析了均值漂移和尺度空间滤波的关系,指出当采用正态核函数时,均值漂移是一种基于高斯尺度空间滤波的聚类方法。由此提出了一种基于均值漂移的非线性尺度空间滤波方法,其性能仅取决于空域的核带宽的选择,在每个采样点处根据信号的局部特征自适应选取幅度域的核带宽。该方法克服了均值漂移滤波存在块状效应的缺点。实验结果表明,该方法的整体性能优于均值漂移滤波、高斯滤波和中值滤波。

To solve the problem of how to choose appropriate kernel bandwidths in mean shift filtering, the relationship between mean shift and scale space filtering was analyzed. The analysis showed that when normal kernel is adopted, mean shift is a clustering method based on Gaussian scale space filtering. On the basis of this analysis, a non-linear scale space filtering method relying on mean shift was presented. The proposed filter was controlled only by kernel bandwidth in spatial domain that was chosen at every sample point adaptively by local signal characteristic. Experiments show that his method overcomes the defect of blocky effect in mean shift filtering, and is superior to mean shift filtering, Gaussian scale space filtering and median filter.