LFTD: A Light and Fast Text Detector

Abstract

Aiming at the problem that the existing text detection technology runs slowly on edge devices and terminal devices with limited storage space and low computing capacity, this paper proposes a method based on A Light and Fast Face Detector for Edge Devices (LFFD) and Connectionist Text Proposal Network. (CTPN) A Light and Fast Text Detector (LFTD). First of all, distinguishing from the current situation of large number of parameters and complex model structure of previous text detectors, this paper is based on the LFFD face detection model. It introduces the characteristics that it does not need to preset a large number of anchor boxes with different sizes and proportions, which makes the detection box network in this paper. The frame is lighter. Secondly, for the problem that the detection range of text and image detection frames is different, this paper improves the label part of LFFD and combines the CTPN method to divide the detection frame of this article into several detection frames according to the font size. The proposed method can theoretically detect Large continuous text scale with 100% coverage. Experiments were performed on popular benchmark datasets (ICDAR11, ICDAR13 and ICDAR15). The proposed method can obtain fast inference speed (NVIDIA TITAN 1080Ti: 131.45 FPS at 640 × 480; NVIDIA 2080Ti at 640 × 480: 136.99 FPS; 640 × 480 Raspberry Pi 3 Type B +: 8.44 FPS), the parameter amount of this model is 8 MB.