Similarity-Based Multi-Target Domain Adaptation for Scene Text Spotting in Inclement Weather Conditions

Inclement weather can cause performance degra-dation on visual tasks. Many single-target domain approaches have achieved significant success in addressing domain adaptation problems in severe weather conditions, however they are designed specifically to handle one type of degradation, and lack a unified framework to handle various inclement weather conditions. In this paper, we propose a method based on a unified framework for unsupervised multi-target domain adaptation to solve the problem of scene text spotting under various inclement weather conditions. The method proposed in this paper aims to dis-tinguish between domain-invariant and domain-specific feature distributions based on sample similarity. We regard similarity samples across different domains as possessing superior domain invariance properties, thus assigning them elevated weights. To achieve this, we introduce the information entropy and Pearson correlation coefficient, to discriminate sample similarity and calibrate the direction of global feature distribution alignment. Experimental results show that our method improved significantly on both synthetic datasets and natural datasets for multiple types of weather.