Improved Maximally Stable Extremal Region detector in color images

Abstract

Maximally Stable Extremal Region (MSER) has been proved to be a powerful local invariant feature. The original MSER detector only utilizes the intensity space. To get more information from a color image, it is naturally to extract MSERs from H, S and I spaces, or other color spaces. However, the increased MSER set inevitably bring in some unstable MSERs, and this will contaminate the over-all reliability. This paper presents an improved MSER detector in color images, which can extract significant increased MSERs, and the results from different spaces are with consistent reliability. In order to eliminate unstable MSERs, mean of the gradient value of region boundaries and the shape index are used as description of the stability of MSER, SVM classifier is used to filter the MSERs from H, S and I space separately. Experimental results demonstrate that the proposed method outperforms the traditional MSER detector.