Zoom Text Detector

Chuang Yang, Mulin Chen, Yuan Yuan, Qi Wang

Abstract

The Zoom Text Detector (ZTD) is a high-speed, high-accuracy framework designed to overcome prevalent challenges in scene text detection, such as feature defocusing and text adhesion. By mimicking a camera's zoom mechanism, ZTD employs a "Zoom Out" module to capture global semantic context and a "Zoom In" module to refine edge details. Additionally, a Sequential Visual Discriminator (SVD) is utilized during training to suppress false positives by leveraging the sequential nature of text. Experimental results demonstrate that ZTD achieves state-of-the-art performance, reaching speeds of up to 97.4 FPS on standard benchmarks while maintaining superior accuracy.

Introduction

Scene text detection is a foundational task for various applications, including autonomous navigation and real-time translation. Current high-performance detectors often utilize "shrink-mask" strategies to prioritize speed. However, these methods frequently struggle with complex backgrounds, leading to feature defocusing, or fail to distinguish between closely spaced text instances, resulting in "text adhesion."

Conceptual overview of the Zoom Text Detector (ZTD) mimicking a camera's zoom process.

The Zoom Text Detector (ZTD) addresses these issues by introducing a multi-scale supervision strategy inspired by a camera lens. By "zooming out" for coarse semantic understanding and "zooming in" for fine-grained boundary detection, the model effectively balances global context with local precision.

Methodology

The ZTD framework consists of three primary novel components integrated into a unified architecture.

The overall architecture of the Zoom Text Detector (ZTD).

Zoom Out Module (ZOM)

The ZOM focuses on coarse-grained training targets. By performing segmentation at a lower resolution, it forces the network to retain high-level semantic information. This prevents the model from becoming overly focused on pixel-level noise, ensuring it can reliably identify text regions within complex scenes.

Zoom In Module (ZIM)

To address the problem of blurry edges and word merging, the ZIM specifically targets the "margins"—the area between the inner shrink-mask and the actual text boundary. By explicitly learning to recognize these boundaries, the model produces sharper contours, effectively preventing nearby text instances from merging into a single detection.

Sequential Visual Discriminator (SVD)

The SVD is a training-only module that treats text as a sequence of features rather than a static texture. It employs a Bidirectional LSTM to analyze the flow of visual features. Since real text possesses an inherent sequential structure that random background textures lack, the SVD effectively filters out visually similar false positives (e.g., fences or grates). Crucially, this module is discarded during inference, adding no computational overhead to the final deployment.

Experiments and Results

ZTD was evaluated on four major benchmarks: MSRA-TD500, Total-Text, CTW1500, and ICDAR2015. These datasets cover a wide range of challenges, including multi-language text, curved shapes, and complex incidental backgrounds.

Performance Metrics

ZTD demonstrates an exceptional balance between speed and accuracy:

  • Speed: On the MSRA-TD500 dataset, ZTD-512 achieved a throughput of 97.4 FPS, significantly outperforming other efficient models like DB (62 FPS) and PAN (30.2 FPS).
  • Accuracy: The model achieved an F-measure of 86.8% on MSRA-TD500 and 84.1% on CTW1500, establishing a new state-of-the-art for high-speed detectors.

Qualitative Analysis

Qualitative results confirm that ZTD's margin-focused approach successfully separates "adhesive" text instances where traditional shrink-mask methods fail. The sequential analysis provided by the SVD also results in a noticeable reduction in false positive detections in texture-heavy environments.

Conclusion

The Zoom Text Detector (ZTD) provides a robust solution to the limitations of current scene text detection frameworks. By integrating multi-scale "zoom" modules and a sequential discriminator, ZTD achieves a superior trade-off between speed and precision. While challenges remain regarding extreme motion blur and low-contrast environments, ZTD represents a significant step forward in real-time, high-accuracy text detection.

Summary updated

BibTeX
@article{yang2024zoom,
  author = {Chuang Yang and Mulin Chen and Yuan Yuan and Qi Wang},
  title = {{Zoom Text Detector}},
  journal = {IEEE Trans. Neural Networks Learn. Syst.},
  year = {2024},
  doi = {10.1109/tnnls.2023.3289327},
}
Figure