Graph-based detection of objects with regular regions

Abstract

Most objects with regular regions could be detected as Maximally Stable Extremal Regions (MSER) [20]. In this paper, We formulate object detection as a bi-label (object and non-object regions) segmentation problem, and propose a graph-based object detection method using edge-enhanced MSER. Specifically, we focus on detecting text in natural images, which is a special kind of object. First, edge-enhanced MSERs are detected as basic letter components; non-text MSERs are then efficiently eliminated by minimizing the cost function which combines both region-based and context-relevant information; and finally, mean-shift clustering is used to group text components into regions. The proposed method is naturally context-relevant, scale-insensitive and readily to be applied on detecting other objects. Experimental results on the ICDAR 2011 competition dataset show that the proposed approach outperforms state-of-the-art methods both in recall and precision. © Springer-Verlag Berlin Heidelberg 2012.