TextFocus: Efficient Multi-scale Detection for Arbitrary Scene Text

Abstract

In the age of deep learning, the emergence of high-resolution datasets containing small text presents a growing challenge in scene text detection. Scaling down entire images to address this issue often leads to text distortion and performance degradation. In this research, we introduce TextFocus, an innovative algorithm that leverages a multi-scale training strategy with a focus on efficiency. Instead of analyzing each pixel in an image pyramid, TextFocus will attempt to identify context regions surrounding ground-truth instances, or “chips,” and then process for finding all text regions in the image sample. All text information from every chip from the model will then be combined with careful post processing methodology to obtain the final results for text detection. As a result of TextFocus’ ability to resample very large image samples (4000x4000 pixels) into low resolution chips (640x640 pixels), our model can train twice as quickly and handle batches as large as 50 on a single GPU when scaled normally. When the larger the training size, the better the result is basic tactic, our method demonstrates that training on high resolution scale might not be ideal. Our implementation using ResNet-18 backbone with segment-like head achieves 0.828 F1 score on the SCUT-CTW1500 [1] dataset, 0.611 F1 score on the Large CTW [2] dataset with acceptable FPS for realtime purpose.