Explicit Relational Reasoning Network for Scene Text Detection

TL;DR

ERRNet tackles the post-processing bottleneck of connected-component based scene text detection by reframing text as an ordered sequence of components and solving detection as a tracking-like problem. It introduces an explicit relational reasoning decoder that models spatial and sequential relationships across components with a Hungarian-matching supervision, enabling end-to-end CC-based predictions. A Polygon Monte-Carlo PIoU is proposed to quantify localization quality, paired with a position-supervised classification loss to align confidence with accuracy. Experiments on CTW1500, Total-Text, ArT, and MSRA-TD500 show ERRNet achieves state-of-the-art or competitive accuracy with strong inference speed, demonstrating the practicality of tracking-based CC representations for efficient scene text detection.

Abstract

Connected component (CC) is a proper text shape representation that aligns with human reading intuition. However, CC-based text detection methods have recently faced a developmental bottleneck that their time-consuming post-processing is difficult to eliminate. To address this issue, we introduce an explicit relational reasoning network (ERRNet) to elegantly model the component relationships without post-processing. Concretely, we first represent each text instance as multiple ordered text components, and then treat these components as objects in sequential movement. In this way, scene text detection can be innovatively viewed as a tracking problem. From this perspective, we design an end-to-end tracking decoder to achieve a CC-based method dispensing with post-processing entirely. Additionally, we observe that there is an inconsistency between classification confidence and localization quality, so we propose a Polygon Monte-Carlo method to quickly and accurately evaluate the localization quality. Based on this, we introduce a position-supervised classification loss to guide the task-aligned learning of ERRNet. Experiments on challenging benchmarks demonstrate the effectiveness of our ERRNet. It consistently achieves state-of-the-art accuracy while holding highly competitive inference speed.

Figures (9)

• Figure 1: (a) Post-processing illustration of typical CC-based methods, which require grouping and ordering components one by one. (b) The pipeline of ERRNet, which has no post-processing. ERRNet views each text instance as multiple text components in sequential movement. Same shape queries indicate predictions in different text instances but in the same sequential position, same color queries represent predictions in one instance, and temporal relationships denote the sequential relationships of components.
• Figure 2: Comparison with leading text detection methods of three different types: (a) Segmentation-based method (TextBPN++), (b) Regression-based method (LRANet), and (c) Connected component-based method (DRRG).
• Figure 3: The architecture of ERRNet, which is mainly composed of three modules: (a) the backbone and feature pyramid network (FPN) for multi-scale feature extraction, (b) the text component initialization module to generate initial component queries, and (c) the explicit relational reasoning decoder for decoding the component sequence for each text instance in order.
• Figure 4: Illustration of ground-truth text component generation for a text primitive. Black point means start point.
• Figure 5: Illustration of our Polygon Monte-Carlo method for calculating the Polygon Intersection over Union (PIoU).