Texture Feature Fusion with Neighborhood Oscillating Tabu Search for High Resolution Image Classification

Multi-channel Gabor filters (MGFs) and Markov random fields (MRFs) are two common methods for texture analysis. This paper investigates their integration through a novel algorithm using the neighborhood-oscillating tabu search (NOTS) for high-resolution image classification. The NOTS algorithm fuses the texture features extracted by MGF and MRF. This algorithm has been compared with classical methods such as sequential forward selection, sequential forward floating selection, and oscillating search. Experimental results show that the fused MGF/MRF features have much higher discrimination than pure features, and NOTS outperforms other algorithms with either pure or fused features. The stability and effectiveness of the proposed algorithm have been verified using Brodatz, Ikonos, and QuickBird images. Introduction With the advent of high-resolution satellite images such as Ikonos and QuickBird, texture analysis has been receiving ever-increasing attention in image classification. Texture reflects the local variability of grey levels in the spatial domain and reveals the information about the object structures in the natural environment. In a high-resolution satellite image, objects such as residential areas and woodlands typically show significant variations in their spectral reflectance. The conventional per-pixel classification methods based upon spectral comparisons are found to be inefficient for classifying such an image with complex textures (Clausi, 2001; Chen and Gong, 2004). Clearly there is a need to incorporate the texture features to improve the classification accuracy. Over the last two decades, many approaches for texture feature extraction have been developed. These include statistical analysis methods, signal processing techniques such as multi-channel Gabor filters (MGFs), stochastic models such as Markov random fields (MRFs), and geometrical methods. Ohanian and Dubes (1992) provide a thorough comparison of these methods. This paper, however, focuses on the use of two common methods: MGFs and MRFs for Texture Feature Fusion with Neighborhood Oscillating Tabu Search for High Resolution Image Classification Liangpei Zhang, Yindi Zhao, Bo Huang, and Pingxiang Li extraction of texture features. MGFs are capable of obtaining multi-scale texture information corresponding to different scales and orientations, while MRFs are capable of capturing the local spatial textural information in an image assuming that the image intensity depends only on the intensities of the neighboring pixels. It is found that the features derived from the above two methods are very different in nature and have low inter-feature correlations (Clausi, 2001). The combined MGF and MRF texture features may provide richer texture information than either of them alone. However, the combined features without selection will produce more dimensions, which may downgrade the performance of the classifiers and might even result in poorer classification accuracy than the original pure feature. In order to reduce data dimensions and improve classification quality, the combination of different features should be processed by feature selection. In this paper, the neighborhood-oscillating tabu search (NOTS) algorithm is proposed to select an optimal feature subset from the pooled set of MGF and MRF features. This algorithm is compared with classical methods such as sequential forward selection, sequential forward floating selection, and oscillating search, using Brodatz, Ikonos, and QuickBird images with individual and fused feature sets, respectively. The experimental results show that NOTS outperforms other algorithms with either pure or fused features. The remainder of the paper is organized as follows. The following section describes the texture feature extraction methods of MGF and MRF followed by an introduction of the NOTS algorithm for solving the MGF/MRF feature fusion problem. The next section discusses three experiments and ends with our conclusions. Texture Feature Extraction Methods Multi-Channel Gabor Filters (MGFs) Each texture can be thought of as containing a narrow range of frequency and orientation components. Texture discrimination can be done by filtering the image with multiple band-pass filters tuned to the dominant frequency and orientation component of the textures. The Gabor filter is a popular method for texture analysis owing to its simplicity PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING March 2008 323 Liangpei Zhang, Yindi Zhao, and Pingxiang Li are with the State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, P.R.China. Bo Huang is with the Department of Geography and Resource Management, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong ([email protected]). Photogrammetric Engineering & Remote Sensing Vol. 74, No. 3, March 2008, pp. 323–331. 0099-1112/08/7403–0 /$3.00/0 © 2008 American Society for Photogrammetry and Remote Sensing 06-031.qxd 2/7/08 8:13 PM Page 323