Abstract
In this paper we propose a new solution to the text detection problem via border learning. Specifically, we make four major contributions: 1) We analyze the insufficiencies of the classic non-text and text settings for text detection. 2) We introduce the border class to the text detection problem for the first time, and validate that the decoding process is largely simplified with the help of text border. 3) We collect and release a new text detection PPT dataset containing 10,692 images with non-text, border, and text annotations. 4) We develop a lightweight (only 0.28M parameters), fully convolutional network (FCN) to effectively learn borders in text images. The results of our extensive experiments show that the proposed solution achieves comparable performance, and often outperforms state-of-the-art approaches on standard benchmarks—even though our solution only requires minimal post-processing to parse a bounding box from a detected text map, while others often require heavy post-processing.
Introduction
Scene text detection in the wild is notoriously difficult because text comes in various fonts, orientations, and scales. Traditionally, researchers have used two main strategies: bottom-up (detecting characters and grouping them) or top-down (detecting blocks and splitting them). Both approaches struggle when text lines are very close to each other. In a standard binary classification setup (text vs. non-text), these adjacent lines often "bleed" together into a single blob in the probability map. To fix this, developers usually resort to "heavy post-processing"—mathematically intense rules to find gaps or split components. This paper aims to move this complexity from the post-processing stage into the learning stage.
Methods: The Border Learning Objective
The core innovation is the introduction of a three-class learning objective: Text (the inner body), Border (a "moat" around each text line), and Non-text (the background). By explicitly teaching the network to find the borders, the model "self-organizes" the text regions. Even if two lines are nearly touching, the learned border class acts as a natural separator.
Architecture
We developed a remarkably lightweight Fully Convolutional Network (FCN) with only 0.28M parameters. We proposed two versions: a single-resolution encoder-decoder and a multi-resolution version that fuses feature maps from four different scales to better handle text of varying sizes. Because the borders effectively isolate text instances, the "decoding" process is reduced to simple connected-components analysis and fitting a bounding box—a "minimal" post-processing step compared to previous state-of-the-art methods.
Datasets
The study utilizes several key benchmarks:
- PPT Dataset (New): 10,692 images from PowerPoint slides, uniquely annotated with three-class labels.
- ICDAR 2015: Standard benchmarks for "born-digital" (web) images and focused scene text.
- MSRA-TD500: A dataset focused on multi-oriented text, including both English and Chinese.
Key Results
- The Power of the Third Class: Adding the border class improved the F-score by a staggering 0.27 on the PPT dataset compared to binary classification.
- Accuracy and Speed: The proposed method achieved an F-score of 0.93 on ICDAR 2015, matching or beating much heavier models. On a GPU, the model processes images at roughly 0.24 seconds per image.
Limitations and Future Work
Standard evaluation metrics often use word-level boxes, whereas our model detects entire lines; this can lead to a mismatch where the model is "penalized" for correctly detecting a line that the metric expects to be split. Future work will explore using learned border shapes to "rectify" (straighten) warped text, which would significantly aid OCR engines in reading text on curved surfaces.
Summary updated