Multiple-level Alignment For Cross-domain Scene Text Detection

Abstract

Deep learning based scene text detection models have achieved promising results on quite a few benchmark datasets. However, most existing methods generally assume that the training and test data are sampled from the same distribution. As a result, in real-world applications where data usually come from various distributions, those models will suffer from performance degradation. In this work, we propose a novel domain adaptation method based on adversarial training to align low-level and high-level features. Specifically, as low-level features focus on the texture information of the image, we propose a so-called Low-level Alignment Module (LAM) to align the low-level features through pixel-level prediction. Meanwhile, we design a High-level Alignment Module (HAM) to align high-level semantic features based on semantic information. We have conducted detailed experiments on two types of domain adaptation tasks in synthetic scenes and natural scenes simultaneously, and the experimental results show the effectiveness of our method.