Reducing the Dimensionality of Feature Vectors for Texture Image Retrieval Based on Wavelet Decomposition

Content-based texture image retrieval based on wavelet decomposition is one of the most active research areas. Subband statistics are normally used to construct feature vectors for calculating the similarity between the example and candidate images. However, most previous methods make no further analysis of the decomposed subbands or simply remove most detail coefficients. The retrieval algorithms commonly use many features without consideration of whether the features are effective for discriminating different classes. This may produce unnecessary computation burden and even decrease the retrieval performance. This paper proposes a method for selecting effective wavelet subbands based on new feature selection functions, which are derived from a modification of fisher´s discriminant. The method can discard those subbands that are redundant or may lead to wrong retrieval results. We test our method using samples from the VisTex texture database, and evaluate the retrieval performances using Daubechies and Gabor wavelet decomposition. The experimental results indicate that, compared with traditional approaches, our method can not only reduce the dimensionality of feature vectors but also improve retrieval performance.