A Prototype System of Content-based Retrieval of Remote Sensing Images

The problem of content-based retrieval of remote sensing images presents a major challenge not only because of the surprisingly increasing volume of images acquired from a wide range of sensors but also because of the complexity of images themselves. In this paper, a prototype software system for content-based retrieval of remote sensing images, namely CBRRSI, is introduced. The main contribution of our research work is the novel wavelet-based feature representation approach and the flexible progressive retrieval strategies based on multi-descriptor. Some experimental results are given to prove the validity of the CBRRSI. . INTRODUCTION Content-based image retrieval (CBIR) has been an active research area in the field of image database and computer vision in recent years. And due to increases in the number of high resolution imaging satellites, available bandwidth and commercial applications of remote sensing, image databases of remote sensing images (RSIs) have been expanding rapidly. So information systems with the ability to storage, browse, index and retrieve information from large databases of RSIs is urgently needed. Although the research on application of CBIR to RSIs is quite not new, the results are far from satisfactory. Currently, CBIR has to face to many open issues including user interface, feature representation, feature indexing structure, similarity measurement and evaluation criterion and the complicated nature of RSIs presents a particular challenge. In this paper, a prototype system for content-based retrieval of RSIs, namely CBRRSI, is introduced. The main contribution of our research work includes the novel wavelet-based feature representation approach and the flexible progressive retrieval strategies based on multi-descriptor. While the color (spectrum) information of a RSI is represented by two descriptors, both of which are obtained from wavelet transform and adopted as coarse filter and detail filter respectively, the texture information is described by fast wavelet histogram and co-occurrence matrix as well to take advantage of both statistics-based and transform-based texture analysis. Based on color and texture feature, flexible progressive retrieval strategies are employed to improve retrieval performance. . RETRIEVAL STRATEGIES A.. Feature Selection and Representation Color and texture are the most meaningful features in RSIs. By taking into account that a RSI is generally visualized as a pseudo-color composition of different spectral bands, before color feature extraction we postulate that all RSIs saved in a database have the same band combinations. HSV color space is chosen for color-based image analysis in our system due to its perceptual characteristic. The color information of a RSI is represented by two descriptors, both of which are generated from the sub-images of a wavelet decomposition of the original image. While the wavelet-coefficient standard deviation of the approximation sub-image of J-level decomposed result is used as a coarse filter, the wavelet coefficient magnitude of the detail sub-images of k-level (k=J,...,1) decomposed results is used as a detail filter. The benefit is obvious. By adopting this progressive strategy retrieval efficiency can be improved significantly and the use of the two descriptors lets our system give attention to both general information and detailed information of a RSI. Besides color, texture plays an important role in the human visual system for recognition and interpretation. Various texture feature extraction techniques such as co-occurrence 0-7803-7930-6/$17.00 (C) 2003 IEEE 0-7803-7929-2/03/$17.00 (C) 2003 IEEE 3700 matrix, MRF, Gabor filters and wavelet transform have been suggested in the past. Among them wavelet transform provides a powerful way for texture analysis due to its multi-resolution property, which is of great significance for RSIs. As a compressed domain indexing technique with good performance based on wavelet transform rather than on pixel domain directly, wavelet histogram technique (WHT) gains popularity in recent years. In our system we adopt fast wavelet histogram techniques (FWHT), which is presented by M.K. Mandal and provides a performance comparable even superior to that of WHT at a reduced complexity [2], and use fast wavelet histogram (FWH) as a primary texture feature index. In addition, co-occurrence matrix features of the original image, approximation and detail sub-bands of 1-level wavelet transform decomposed images are calculated in order to improve retrieval performance. By combining wavelet statistical features and co-occurrence features our system takes advantage of both statistics-based and transform-based texture analysis. B. Progressive Retrieval Strategies In our system the retrieval strategies are flexible. Users can adopt different retrieval strategy in different case, i.e. users can select any one of the two representations (color and texture) or both to index a RSI. If both are desired, users can decide which one has top-priority or both have the same priority with different weight. In the next part we will introduce our progressive retrieval strategies based on color and texture respectively. The first step of color-based retrieval is color space conversion from RGB to HSV. Then each color component is represented by a set of wavelet-based decomposed sub-images at several scales including three orientation selective detail images and a coarse or approximate image. In our system the maximum decomposition level is three (J=3). And then wavelet-coefficient standard deviation of the approximation sub-band of J-level decomposed images are calculated as an index, which is used as a coarse color filter to quickly screen non-promising images from further consideration using (1), where t LL i , σ and qLL i , σ represent wavelet-coefficient standard deviation of each color component of the J-level approximation sub-image of target image and query image respectively, TV i D represents the threshold value of similarity distance and i represents H, S, V component respectively. TV V q LL V t LL V TV S q LL S t LL S TV H q LL H t LL H D D D ≤ − ≤ − ≤ − , , , , , , & & & & σ σ σ σ σ σ (1) The detail color filter, described by the weighed result of wavelet coefficient magnitude obtained from detail sub-images of k-level (k=J,...,1) decomposed results, is subsequently used to further search the desired images using (2), where k represents the LH, HL and HH directional information, α and β represent the weight of directional information and the weight of color component respectively, and x represents the wavelet coefficient magnitude. ( ) TV k t k V q k V t k S q k S t k H q k H