Text detection is the basis of Optical Character Recognition (OCR) and text information retrieval from natural images.The intrinsic variability of text renders scene text detection more challenging than traditional object detection.This paper presents a new coarse-to-fine approach which aims at improving scene text detection in two aspects.First of all, a novel Double-threshold Gradient Patterns (DGP)feature, which shows robust invariance to background contrast reversal and can adaptively eliminate noise interference in scene images, is proposed to depict the scene text texture.Secondly, by virtue of the emergent Extreme Learning Machine (ELM), we propose a cascade weighted ELM detector to discriminate text from non-text with almost minimal training time, which makes the proposed algorithm more applicable and scalable.Experiments on ICDAR 2003 dataset demonstrate the effectiveness of our approach in comparison with existing methods.