On Exploring and Improving Robustness of Scene Text Detection Models

TL;DR

This paper evaluates the robustness of scene text detection models against corruptions using new datasets, analyzes key components affecting robustness, and proposes a simple method to enhance model resilience.

Contribution

It introduces two new datasets for robustness evaluation, systematically analyzes model components, and proposes a data-based method to improve robustness of scene text detection models.

Findings

Robustness varies significantly across different model components.

The proposed background-foreground merging method improves robustness.

New datasets enable better evaluation of robustness in scene text detection.

Abstract

It is crucial to understand the robustness of text detection models with regard to extensive corruptions, since scene text detection techniques have many practical applications. For systematically exploring this problem, we propose two datasets from which to evaluate scene text detection models: ICDAR2015-C (IC15-C) and CTW1500-C (CTW-C). Our study extends the investigation of the performance and robustness of the proposed region proposal, regression and segmentation-based scene text detection frameworks. Furthermore, we perform a robustness analysis of six key components: pre-training data, backbone, feature fusion module, multi-scale predictions, representation of text instances and loss function. Finally, we present a simple yet effective data-based method to destroy the smoothness of text regions by merging background and foreground, which can significantly increase the robustness of different text detection networks. We hope that this study will provide valid data points as well as experience for future research. Benchmark, code and data will be made available at \url{https://github.com/wushilian/robust-scene-text-detection-benchmark}.