Scene Text Detection with Adaptively Weighted Descriptors in Opponent Color Space

Abstract

This paper involves the challenging problem of text localization in complex scenes. Maximally Stable Extremal Regions (MSER) has recently been extensively employed in text detection methods. However, most efforts for MSERs are only put into a single intensity image. MSERs are therefore subjective to cases like low contrast and uneven lighting etc. In this paper, we investigate the opponent color theory and perform attempting to evaluate the benefit of MSER extraction on different opponent color channels. Furthermore, we propose to apply kernel descriptors for text classification and multi-kernel learning to learn relative weights for features. We have experimented our proposed strategy on ‘Robust Reading Competition’ dataset distributed by International Conference on Document Analysis and Recognition (ICDAR) 2003 and 2011. The experiment results demonstrate its effectiveness and efficiency. We can achieve general equivalently good performance with several compared state-of-the-art methods at lower computation cost.