|
Patch clustering is a common issue in image processing and pattern recognition, especially in those patch- based structured sparsity reconstruction problems. Researchers usually adopt the K-means method based on the gray intensity distance or partition according to the edge direction. However, these metrics are not sufficient to help obtaining delicate classification. In this letter, we propose a novel image patch clustering method based on the magnitude spectrum structure and directionality in Fourier domain (SSDF), i.e. the primary direction in the spectrogram, the spectrum structure complexity and components distribution of low, middle and high frequency. Experimental results demonstrate that SSDF method is able to achieve more exquisite classification following three steps of subdivisions with no need to preset the cluster number. |
|
Keywords:Image processing; patch clustering; structured sparsity; magnitude spectrum in Fourier domain; primary direction; structure complexity |
|