New wavelet model for texture-based feature extraction and retrieval

Abstract

As a picture can describe a thousand of words, it is necessary to develop a computer management system that can efficiently handle image/video data. Subsequently, a good image/video database system should provide a fast and accurate method that can help us handle the matching problem of visual features. Texture can be recognized as an important feature for content-based image/video feature similarity matching. One of the objectives of this research is to investigate a fast and accurate texture feature retrieval method which can be efficiently used on an image/video database for similarity searching. We suggest using the over-complete wavelet-based model for texture feature analysis. The method matches the viewpoint of human vision on multi-channels, frequencies and orientations of texture properties. The computation time for this texture analysis is small as compared to other well-known texture analysis methods, such as the Gabor wavelets. A new texture analysis approach from the Laplacian of Gaussian(LOG)-based over- complete wavelets is derived. It is proven to be suitable for texture analysis from both theoretical and experimental studies. Also, a new texture representation that can significantly improve the retrieval accuracy of texture features for both over-complete and sub-sampling wavelet schemes is given. We have compared different texture-based feature retrieval methods, and experimental results indicate that our proposed method achieve the highest retrieval rate on the entire Brodatz texture database with a low feature analysing time. It is thus suitable for real-time content-based retrieval applications. We have also analysed the retrieval performance of various texture-based feature retrieval methods under different Gaussian noise levels. Results of the study indicate that the over-complete wavelet scheme is robust against noise, due to its translation invariant property. The traditional over-complete wavelet implementation contains redundancy between filtering structures. Also, it causes boundary artifact in reconstructed images. Therefore, based on LOG-base over-complete wavelet transform derived for texture analysis, we have investigated a fast implementation method which aims to reduce the complexity and boundary artifact of wavelet transforms. Results of our theoretical and experimental studies show that the computational complexity can be reduced significantly by using the spatial approach as compared to the filtering approach. Furthermore, it can effectively eliminate the boundary artifacts, during multi-resolution wavelet transforms effectively, and thus perfect reconstruction is always obtained.