A Robust Blob Detection and Delineation Method Liang Wang, Hehua Ju School of Electronics Information and Control Engineering, Beijing University of Technology, Beijing, China {wangliang, juhehua}@bjut.edu.cn Abstract—This work presents a robust method to detect blob and fit its contour in image. Previous methods for blob detection and delineation were either liable to fail with outliers and noise or computationally expensive. By incorporating the prior information of the region-of-interest and introducing the concept of the kernel MSER, the modified MSER detection method can detect the unique blob which is the most stable region to represent the blob. For further processing, the constrained least squares method by incorporating pruning technique is used to fit the ellipse corresponding to the contour of the detected blob. With this proposed method, we can detect and fit blob with high accuracy. The experiments show the validity of the proposed method. Keywords-blob detection; blob delineation; kernal MSER; the constrained least squares method; ellipse fitting I. INTRODUCTION Blob is the region that is either brighter or darker than the surrounding, or in the same color in the image or video. Object, mark and human in videos and images can be modeled as blob. Detection and delineation of blob are fundamental research task in computer vision. It can be widely used in visual surveillance, motion capture, object tracking and robot’s visual localization and navigation. Many works [1,2,3,4,5] have been done in blob detection and delineation. For blob detection, there are mainly two classes of methods: (i) differential methods based on derivative expressions and (ii) methods based on local extrema in the intensity landscape. For the former, the method based on the Laplacian of the Gaussian (LoG) is the first and also most common one [1]. The strong positive responses for dark blobs and strong negative responses for bright blobs can be obtained with this method. So blob can be detected by analyzing the responses. A main problem is that the response is strongly dependent on the relationship between the size of the blob structures in the image domain and the size of the Gaussian kernel used for pre-smoothing. Therefore some multi-scale approaches are proposed to capture blobs of different (unknown) size. And some modifications, such as using Difference of Gaussians (DoG) or the determinant of the Hessian instead of LoG, are adopted to prove performance [2,3]. However the computation is still expensive and the proper scale of blob corresponding to the interesting object is hard to automatically obtain. For the latter, they associate a bright (dark) blob with each local maximum (minimum) in the intensity landscape. Some image processing methods are used to determine the interesting region [4,5,6]. A main problem with such these approaches, however, is that local extrema are very sensitive to noise and outlier. Once the blob being detected, descriptor or model fitting is needed to delineate the blob for further processing. For most applications, the contour of the blob may be modeled as an ellipse with ellipse fitting. The literature on ellipse fitting divides into two classes: clustering (such as Hough-based methods)[7,8] and least squares fitting [9,10]. Clustering techniques center on detecting the maximum in the parameter space by voting. But it has expensive computation cost. Least-squares techniques concentrate on finding the set of parameters to minimize some distance measures between the data points and the ellipse. But noise and outliers may easily spoil the fitting result. A robust blob detection and model fitting method is proposed in this paper. A modified maximally stable extremum region (MSER) method is used to detect blob to overcome the effect of noise and outlier. Then a constrained least squares method incorporating pruning technique is used to fit ellipse to represent the detected blob. The rest of this paper is organized as follows. In Section 2 blob detection method based on MSER is given in detail. In Section 3, a robust ellipse fitting method for blob contour is presented. Experiments are reported in Section 4. Section 5 presents some conclusions. II. BLOB DETECTION WITH MODIFIED MSER The concept of maximally stable extremum region (MSER) was originally defined in [6], which was used to establish tentative correspondences between a pair of images taken from different viewpoints. Its principle can be explained informally as follows. Imagine all possible thresholds of a gray-level picture I. The pixels below a threshold are regarded as ’black’, and those above or equal to that are regarded as ’white’. Then we have: 1 () () 0 t if Ix t E x otherwise > ⎧ = ⎨ ⎩ (1) If we increase the threshold t continually, a movie of binary pictures t E with frame t corresponding to threshold t can be obtained. Firstly a white frame would be seen. Then with the increasing of threshold black spots corresponding to local intensity minima will appear and grow. And regions corresponding to two local minima will merge and become a larger black region. Finally, the last frame will be black. Maximal regions are those connected components of all frames of the movie. Similarly minimal regions could be obtained by inverting 978-0-7695-3563-0/08 $25.00 © 2008 IEEE DOI 10.1109/ETTandGRS.2008.294 827 2008 International Workshop on Education Technology and Training & 2008 International Workshop on Geoscience and Remote Sensing 978-0-7695-3563-0/08 $25.00 © 2008 IEEE DOI 10.1109/ETTandGRS.2008.294 827