A Modified PSO Algorithm for Remote Sensing Image Template Matching

Image template matching is essential in image analysis and computer vision tasks. Cross-correlation algorithms are often used in practice, but they are sensitive to nonlinear changes in image intensity and random noise, and are computationally expensive. In this paper, we propose a templatematching algorithm based on a modified particle swarm optimization (PSO) procedure with a mutual information (MI) similarity measure. The influence of MI on the performance of template matching, calculated by different histogram bins, is analyzed first. A modified PSO method (CRI-PSO) is then presented. The proposed algorithm is tested with remote sensing imagery from different sensors and for different seasons. Our experimental results indicate that the proposed approach is robust in practical scenarios and outperforms the standard PSO, multi-start PSO, and cross-correlation algorithms in accuracy and efficiency with our test data. The proposed method can be used for position estimation of aircraft, object recognition, and image retrieval. Introduction Image template matching is a process of determining the presence and location of an input image or an object inside a reference image (Choi et al., 2002). The input image and the reference image may be dissimilar because they may exhibit relative translation, rotation, and different scales or contain additive random noise due to different sensors and image captured at different dates. All these factors make PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Ap r i l 2010 1 Ru An, Qing Zhang, Chunye Chen and Peng Yan are with the Department of Geographical Information Sciences, College of Hydrology and Water Resources, Hohai University, Najing 210098, China ([email protected]). Peng Gong is with the Department of Environmental Science, Policy and Management, University of California at Berkeley, 137 Mulford Hall, Berkeley, CA 94720, and State Key Laboratory of Remote Sensing Science, Institute of Remote Sensing Applications of Chinese Academy of Sciences and Beijing Normal University, Beijing 100101, China. Huilin Wang, Xuezhi Feng, and Pengfeng Xiao are with the Institute of Geography and Sea Sciences, Nanjing University, Nanjing 210093, China. Qi Chen is with the Department of Geography, University of Hawai’i at Manoa, 422 Saunders Hall, 2424 Maile Way, Honolulu, HI 96822. Qing Zhang, Chunye Chen and Peng Yan are with the Department of Geographical Information Sciences, College of Hydrology and Water Resources, Hohai University, Najing 210098, China. Photogrammetric Engineering & Remote Sensing Vol. 76, No. 4, April 2010, pp. 000–000. 0099-1112/10/7604–0000/$3.00/0 © 2010 American Society for Photogrammetry and Remote Sensing A Modified PSO Algorithm for Remote Sensing Image Template Matching Ru An, Peng Gong, Huilin Wang, Xuezhi Feng, Pengfeng Xiao, Qi Chen, Qing Zhang, Chunye Chen, and Peng Yan image matching a challenging task (Brown, 1992; Zitova et al., 2003). Image template matching has been widely used for object recognition, image retrieval, motion tracking, aircraft position estimation, change detection, and multiimage registration (An et al., 2005; Gong et al., 1992; Liu et al., 2006; Oh et al., 2006; Sim et al., 1999a and 1999b). Cross-correlation-based algorithms such as the normalized cross-correlation (NCC), the mean absolute error (MAE), and the mean squares error (MSE) methods are often used in image template matching. These methods do not generally require extensive preprocessing, such as segmentation or feature extraction, but they often have a lower probability of correct matching when the image intensity has nonlinear changes. They are also computationally expensive. Other commonly used methods in template matching are featurebased techniques. These methods often fail when the images to be matched contain few salient features such as points, lines, and regions (Brown, 1992; Zitova et al., 2003). Position estimation of aircraft through image matching is an important method for autonomous navigation. Because of its benefits of higher self-determination and accuracy, the development of image-based navigation systems has been a hot research topic (Sim et al., 1999a and 1999b; Oh et al., 2006). Various approaches, such as terrain contour matching (TERCOM), inertial navigation systems (INS), and global positioning systems (GPS), have been used for navigation, but they all have some drawbacks. For example, the estimation error by the INS tends to increase as an aircraft goes on flying, the GPS can be disturbed by external signals, and terrain contour matching encounters difficulties in estimating its own position in the plain regions where change in the elevation is small and can be out of control by external signals (Oh et al., 2006). In practice, a hybrid system, such as, INS and image-based navigation system is commonly used. INS usually supplies orientation and altitude information of the aircraft and image-based navigation is used to rectify cumulative error and increase accuracy of INS systems. In general, navigation systems have been developed under the assumption that an aircraft flies along a predetermined (planned) trajectory. Therefore, it is assumed that the system contains reference images, and it is thus possible to measure the performance of various matching methods (Oh et al., 2006). A matching algorithm suitable for position estimation of aircraft has to meet the requirements of reliability (higher success rate for matching), accuracy