Learning oriented dictionary for sparse image representation

A novel oriented dictionary is proposed for sparse image representation. The scheme of dictionary learning combines the double sparsity model and the zero-tree structure in the wavelet domain. The dictionary atoms are constructed by grouping the wavelet bases in all high-frequency subbands of the same orientation across different scales. This scheme overcomes the limit on the input signal dimension as well as the over-fitting problem. We demonstrate the potential of the proposed dictionary for M-term approximation of fingerprint images.