Texture classification method via random feature dictionary.

Extracting global texture feature through sparse representation faced some problems, which mainly caused by high dimension. In order to solve those problems, this paper proposed a feature extraction and classification method based on random feature dictionary. The proposed method utilized the distribution of non-zero coefficients, which were computed by sparse decomposition, to generate a statistics histogram feature. The acquired histogram could reflect the dictionary atoms' using frequency in sparse decomposition, and was able to reflect the class information. Thus, the classification could be realized. For the sake of improving classification accuracy, it fused multi-scale and multi-direction wavelet features through random projection, and then trained a more descriptive dictionary by those fused features. In the classification experiments, it achieved 94.79% classification accuracy. Further experiments and analysis prove that the proposed method is robust under noisy and bad illumination conditions, and has characteristics of effective and stable in global texture feature extraction. [ABSTRACT FROM AUTHOR]