A model-based method for rotation invariant texture classification
IEEE Transactions on Pattern Analysis and Machine Intelligence
Digital image processing
| Hi-index | 0.14 |
A unified feature extraction scheme, the two-dimensional (2-D) linear prediction model-based decorrelation method, is presented. By applying 2-D causal linear prediction model to decorrelate a textured image, the very heavy computation load required when using a whitening operator to decorrelate the image, or the significant information loss when using the gradient operator to approximately whiten the image is avoided. The texture model-based decorrelation provides three sets of features to perform texture classification: the coefficients of the 2-D linear prediction, the moments of error residuals and the autocorrelation values. An optimum feature-selection scheme using modified branch-and-bound method was introduced to reduce information redundancy. After feature selection, 100% classification accuracy was achieved for a 20-class texture problem. Experiments show that this feature extraction scheme is truly information lossless, effective, and fast.