A Multi-Oriented Scene Text Detection Method Based on Location-Sensitive Segmentation

Abstract

In recent years, regression-based scene text detection methods have achieved great success. However, because the network has a limited receptive field, the predicted bounding boxes cannot enclose the entire text instance when dealing with the long text instance. In this paper, we propose a multi-oriented scene text detection method based on location-sensitive segmentation. The main idea is that we divide the whole text instance detection into three sub-text instances (left part, middle part, and right part) detection. To form the final detection bounding box, we get three candidate bounding boxes from three sub-text instances and then merge them by getting the minimum rectangular area. Finally, the pixel-level score maps are used to filter false positives. Experiments on ICDAR2015 and MSRA-TD500 demonstrate that the proposed method achieves great performance. For ICDAR2015 Dataset, the method achieves an F-measure of 0.822 and a precision rate of 0.876.