Two-dimensional local discriminant basis algorithm for texture classification

Local discriminant basis (LDB) is known as a powerful and efficient feature extractor for signal classification. As LDB is based on wavelet transform it can effectively represent a nonstationary signal with a few number of significant coefficients. On the other hand the redundant orthogonal basis which comes with wavelet packet transform, makes it quite attractive for classification. In this paper the LDB method is extended to two-dimension and applied to texture classification problem. For this purpose three 256 /spl times/ 256 texture images from Brodatz (1968) are selected. Each Image is divided to 2 equal subimages. From each half 64 /spl times/ 64 subimages are selected after randomly finding a top-left point pixel. The miss-classification rate of 20%-30% is obtained.