MDDA: A Multi-scene Recognition Model with Multi-dimensional Domain Adaptation

Abstract

Optical Character Recognition is widely used in scenarios from license plate recognition, chip detection to text translation. However, in some special scenarios, there are very few data for training or the target datasets are not marked. This paper mainly aims to establish a multi-scene text recognition model that can achieve accurate results even when the target test dataset is not marked. The model reduces the distance between domains at the feature level by using an adversarial loss and flipping the gradient. For the adversarial loss, each letter has a different domain discriminator to align the features at the character level. We also align the source domain features and target domain features in the background. Experimental results show that our model achieves good performance when the training dataset and test dataset have domain shift, and suggest possible scenarios for potential applications.