A Multi-scale Domain Adaptive Framework for Scene Text Detection
Wang Bi, Yu Gu, Yu Leo Lu · 2023
Scene text detection is a key and challenging research task in the area of computer vision, which has achieved superior performance with the development of deep learning. However, the model which had highly results relies on large-scale labeled datasets generally as well as large-scale manual labeling was laboriously and difficultly. Due to this, there is indeed a large difference between train and test data which may lead to a severe performance drop. To tackle this issue, we propose a domain adaptation model based on semantic segmentation network aiming at improving cross-domain robustness for scene text detection. Specifically, our model improving the performance of cross-domain problem on two components:(1) global-feature domain classifier module and (2) image-level domain classifier module, which can reduce the feature distributions and the scale difference of source domain and target domain effectively, and we add a consistency regularizer between the two modules to improve the generalization ability of the model. We set up two cross-domain scenarios to evaluate our proposed method. Empirical results demonstrate the effectiveness of our proposed model for domain adaptation scene text detection.