Abstract
The importance of texture analysis is revealed in many applications and filtering methods are demonstrated by experiments to have many vision applications. Thus motivated our study on further investigating filtering methods for texture analysis, in particular, in texture classification and texture segmentation. First, we study the Gabor filter for it simulating human visual system. Traditional Gabor filter faces a problem of heavy computations. We propose a new method, called the multi-resolutional Gabor filter, for reducing the computation time. By merging the multi-resolution concept with the traditional Gabor filter, we only need 42% computation time but preserve the recognition rate. We also propose a new texture segmentation method based on the Gabor filter. Experiments using the proposed subimage binarization and subimage combination method achieve the segmentation results that close to a visual judgement of texture regions. In addition to the non-orthogonal Gabor filter, we also study the wavelet transform for its benefits on orthonormal bases, orientation selectivity, and multi-resolution decomposition. A heritage of wavelet coefficients called truncated wavelet coefficients is proposed for texture classification. Experimental results show that the proposed truncated wavelet features achieve better recognition rate than traditional wavelet features obtain. The wavelet coefficients are also applied for texture segmentation. The proposed approach consists of two major steps: subimage segmentation and subimage combination. In the first step, a mixed texture is decomposed to subimages according to different scales and each subimage is then segmented to a certain number of regions by avoiding over segmentation. Then, a nearest-neighbor subimage combination approach is applied to combine all segmentation results together. Experimental results show that the proposed approach is efficient of segmenting an image consisting of various texture regions. At last, some commonly used features, derived from filtering methods such as Fourier transform, spatial filtering, Gabor transform, wavelet transform, and our proposed features are compared by experiments for texture classification and segmentation. Experimental results show that the proposed multi-resolutional Gabor features are preferred with few number of features, say, 2 to 7, whereas, the truncated wavelet features by zerotree wavelet transform are suggested when more than seven features are requested. Experiments also suggest that our proposed 2-step approach may be used for texture segmentation.