Multi-lingual Scene Text Detection Based On Fully Convolutional Networks

Among the tasks of computer vision,text detection in natural scenes is a challenging task.Text detecting in natural scenes can help the computer to understand and describe the scene.The cognitive ability of the computer in the environment is basic ability of auto driving car,home robots and other automation equipment.With the accumulation of multimedia data on the Internet and the improvement of GPU computing capability,deep learning technology has been widely used in various tasks of computer vision.In some tasks,the performance of deep learning far exceeds that of some traditional algorithm.In the subtitle detection of multimedia video,the traditional methods,such as color or spatial scale and text proportion,have achieved good results.However,the traditional algorithm has a bad result on the multi-lingual text in natural scenes.In this paper,we propose a highly efficient and feasible method,by verifying the generalization ability of single-language text detection models,the single language detection model is migrated to the task of multi-language text detection with transfer learning.Deep learning methods require a large number of labeled training data.In data augmentation,the data generation system can generate natural scene data with multi-scale glyphs and rich fonts,and solves the problem of training data.In text detection algorithm,the semantic segmentation model is migrated to the tasks of text detection.The single language character data is used to train the VGG model to get single-language character classifier,and the VGG is migrated to the full convolutional network as Feature extraction.In terms of model generalization,the single language detection model is applied to the task of multilingual text detection through the transfer learning,and the single language detection model is fine-tuned by the multi-language dataset.The single-language model is verified to have the ability to detect multi-lingual text in natural scenes.The deviation of single language detection model and multil-ingual detection model shows that the single-language detection model has the ability to detect multi-lingual text in natural scenes.