Pyrboxes: An efficient multi-scale scene text detector with feature pyramids

摘要

Scene text detection has attracted many researches due to its importance to various applications. However, current approaches could not keep a good balance between accuracy and speed, i.e., a high-performance accuracy but with a low processing speed, or vice-versa. In this paper, we propose a novel model, named PyrBoxes, for efficient and effective multi-scale scene text detection. PyrBoxes consists of an SSD-based backbone that utilizes deep layers with strong semantics to detect texts in various sizes, and a proposed grouped pyramid module that leverages basic layers to append detailed locations into detection. Most existing detectors discard features from the basic layers due to the efficiency issue. We argue these layers contain fine-grained information, which is complementary to high-level semantics. Based on this, the grouped pyramid module combines the basic layers recursively into a detection layer via a top-down partition and a bottom-up group. Extensive experiments on both horizontal and oriented benchmarks, including ICDAR2013 Focused Scene Text, ICDAR2015 Incidental Text and COCO-Text, demonstrate that PyrBoxes achieves state-of-the-art or highly competitive performance compared with baselines, while runs significantly faster at inference. Furthermore, by experimenting on another ChiTVText dataset, PyrBoxes shows great generality to Chinese and long text lines. By visualizing some qualitative results, as expected, PyrBoxes provides more accurate locations and reduces the rate of missed detections, especially for small-sized texts. (C) 2019 Elsevier B.V. All rights reserved.