BIT-PLANE EXTENSION TO A CLASS OF INTENSITY-BASED CORNER DETECTION ALGORITHMS Ambar Dutta, Animesh Mandal, B. N. Chatterji and Avijit Kar Department of Computer Engineering, B.I.T., Mesra at Kolkata, 700107, Kolkata, India Department of Computer Science & Engineering, Jadavpur University, 700032, Kolkata, India Department of Computer Science & Engineering, B.P.P.I.M.T., 700052, Kolkata, India Department of Computer Science & Engineering, Jadavpur University, 700032, Kolkata, India email: adatta@bitmesra.ac.in , mandal.ani@gmail.com , bnchatterji@gmail.com , avijit_kar@cse.jdvu.ac.in ABSTRACT Corners are important image features whose detection is very important in many computer vision tasks. In this paper we have evaluated the performance of five intensity based corner detectors with the help of a dozen test images, artificial and real, based on six performance measures of which three are proposed by us. We then propose a new approach, using bit- plane decomposition, in which a grayscale image is first divided into several bit-planes, and then the original corner detectors are applied on all the bit-planes simultaneously and finally using a threshold, all the higher bit-plane corners are recombined up to the thresholded bit-plane to obtain the final set of corners. Using this approach, we have seen that the performance of the algorithms has improved significantly with respect to both detection and time. 1. INTRODUCTION Corners in an image are those points where a strong two- dimensional intensity change has been observed in all directions. So, corner detection is regarded as a very important research area for many applications in computer vision such as scene analysis, image registration, image matching, object recognition etc. Since 1977 vision researchers developed a number of corner detection algorithms. The performance of a corner detection algorithm depends on the four factors – accuracy, consistency, efficiency with respect to time and robustness with respect to noise. Zheng et al [14] and Mokhtarian et al [9] provided are good literature survey of the existing corner detectors. Corner detection algorithms can be roughly divided into two categories: boundary based and intensity based. In this paper we restrict ourselves only to intensity-based methods. Moravec [10] detected corners at the locations where a significantly high intensity variation is found in all directions. Kitchen and Rosenfeld [6] derived a cornerness measure by applying differential operators consisting of first and second order partial derivatives of an image to detect corners. Harris and Stephens [5] estimated the cornerness measure based on local autocorrelation using first order derivatives. Deriche and Geraudon [7] detected corners using a scale-space based approach that combined important properties from Laplacian and Beaudet’s [2] cornerness measures. Wang and Brady [13] proposed a corner detection algorithm based on the cornerness measurement of the total surface curvature. Smith and Brady [11] proposed SUSAN corner detector using the concept that each image point is associated with a local area of similar brightness. Laganiere [7] presented a morphological corner detector. Trajkovic and Hedley [12] presented a fast corner detector using a mutigrid approach. Zheng, Wang and Toeh [14] presented an extended version of Plessey corner detector. Golightly and Jones [4] presented an algorithm for both corner detection and matching for visual tracking of power line inspection. Mikolajczyk and Schmid [8] proposed a scale and affine invariant corner detector. Alkaabi and Deravi [1] presented a fast corner detection algorithm based on pruning candidate corners. In this paper we have considered five intensity based methods proposed by Moravec (A1), Kitchen and Rosenfeld (A2), Harris and Stephens (A3), Laganiere (A4) and Alkaabi and Deravi (A5) and observed their performances on a set of dozen images on the basis of six performance measures of which three are proposed by us and three are taken from the literature [4, 9]. We have next proposed an extension to each of these five algorithms using bit-plane decomposition and observed from the graphs that the performance of each of the algorithms has significantly improved. The algorithms BPA1, BPA2, BPA3, BPA4 and BPA5 are the bit-plane extensions to A1, A2, A3, A4 and A5 respectively as referred in the graphs (Graph – 1 to Graph – 4). This paper is organized as follows. In Section 2, we briefly describe the bit-plane decomposition technique. In Section 3, we discuss our proposed approach. Experimental results are then presented and compared with the original version of the algorithms in Section 4. We, finally, draw our conclusions in Section 5. ©2007 EURASIP 267 15th European Signal Processing Conference (EUSIPCO 2007), Poznan, Poland, September 3-7, 2007, copyright by EURASIP