A Filtering Strategy for Interest Point Detecting to Improve Repeatability and Information Content

This paper compares several stereo image interest point detectors with respect to their repeatability and information content through experimental analysis. The Harris-Laplace detector gives better results than other detectors in areas of good texture; however, in areas of poor texture, the HarrisLaplace detector may be not the best choice. A featurerelated filtering strategy is designed for the Harris-Laplace detector (as well as the standard Harris detector) to improve the repeatability and information content for imagery with both good and poor texture: (a) the local information entropy is computed to describe the local feature of the image; and (b) the redundant interest points are filtered according to the interest strength and the local information entropy. After the filtering process, the repeatability and information content of the final interest points are improved, and the mismatching then can be reduced. This conclusion is supported by experimental analysis with actual stereo images. Introduction Image feature extraction plays an important role in the field of image matching, object description, movement estimating, and object tracking. Interest points are the essential elements of an image feature, where an interest point simply means any distinctive point in the image for which the signal changes two-dimensionally. In image analysis and stereovision, the choice of an interest point detector to deal with different application requirements is especially significant (Schmid et al., 2000; Sebe et al., 2003). For the purpose of stereo image matching and the subsequent three-dimensional (3D) reconstruction, only the detecting efficiency and location accuracy were considered in detail in the past when choosing an interest point detector (Brand and Mohr, 1994; Baker and Nayar, 1999). Schmid et al. (2000) pointed out that the choice of an interest point detector should be based on its repeatability and information content. The repeatability of interest points determines the matching reliability, while the information content indicates the significance of such interest points to the 3D object reconstruction. A Filtering Strategy for Interest Point Detecting to Improve Repeatability and Information Content Qing Zhu, Bo Wu, and Neng Wan In experiments with general images, the Harris detector gives better repeatability and information content than other detectors (Schmid et al., 2000). However, for large-scale aerial images or satellite images, many poor interest points would be detected by the Harris detector in imagery with poor texture, and some interest points with small interest strength (also called interest-values, calculated through the response formulation given in Harris (1988)) decrease the repeatability and information content, increase the probability of mismatch, and lower the efficiency of the subsequent image matching. Although some improved methods to the Harris detector such as Mikolajczk and Schmid (2004) proposed to strengthen its invariance to scale and affine transformations, the interest points detection with good repeatability and information content for the stereo image matching and the 3D reconstruction is not studied thoroughly. This paper proposes a filtering method related to the local image texture features to pick out those points that will decrease the overall repeatability and information content. This method takes into account both the interest strength and the local feature of interest point at the same time. A filtering formulation is presented to calculate the response of every pixel, and then a threshold is designed for selecting the interest points. This paper makes two contributions. First, in the next section the concepts of repeatability and information content as applied to stereo image interest point detection are outlined, and the following section compares several typical detectors, which include the traditional and the up-to-date methods, and presents the comparison results in repeatability and information content using standard test images. The second contribution is by making use of the image feature analysis based on information entropy; a filtering method to select interest points related to the local image feature is introduced. The final section describes the detail of this method, and illustrates the improvement of its repeatability and information content through experimental analysis. Repeatability and Information Content Repeatability Given a 3D point P and two projection matrices M1 and M2, the projections of P into two images I1 and I2 are p1 M1P and p2 M2P. The point p1 detected in image I1 is repeated in image I2 if the corresponding point p2 is detected in image I2. To measure the repeatability, a unique relationship PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING May 2007 547 Qing Zhu and Neng Wan are with the State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, 129 LuoYu Road, Wuhan, 430079, P.R. China. Bo Wu was with the State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, and is currently with the Mapping and GIS Laboratory, Department of Civil and Environmental Engineering and Geodetic Science, The Ohio State University, 470 Hitchcock Hall, 2070 Neil Avenue, Columbus, OH 43210 ([email protected]). Photogrammetric Engineering & Remote Sensing Vol. 73, No. 5, May 2007, pp. 547–553. 0099-1112/07/7305–0547/$3.00/0 © 2007 American Society for Photogrammetry and Remote Sensing 05-052 8/6/07 6:26 PM Page 547