| With the development of deep learning,scene text detection has become one of the research hotspots of computer vision.Since natural scene text detection can extract meaningful text information from images of real scenes,it has aroused the increasing interest of researchers and has achieved fruitful research results.However,due to the lack of a well-labeled scene Mongolian image dataset in this field,the research on Mongolian text detection in complex scene images is still in the exploratory stage.At the same time,scene text detection faces problems such as large changes in text scale,insufficient image illumination or insufficient focus,etc.On this basis,the main research contents of this article are as follows:1.This thesis constructs a scene Mongolian dataset.With the continuous popularization of information and the continuous development of Mongolian culture,a large number of Mongolian documents and natural scene text images have been produced.A lot of predecessor research work has been accumulated on the detection of Mongolian documents and documents,but there are relatively few studies on the detection and recognition of Mongolian characters in the scene.There are two main reasons: On the one hand,the research on Mongolian scenes has been carried out late.On the other hand,due to the lack of real standardized training samples,it is difficult to apply the deep learning method of scene text detection to the field of scene Mongolian research.This research proposes an annotated Mongolian scene text dataset.The dataset is taken from real nature scene,such as billboards,traffic signs,school corridors and other scenes.2.This thesis proposes a scene Mongolian text detection method based on improved EAST.Firstly,aiming at the defect that the EAST algorithm is not sensitive to multi-scale text detection,this paper proposes a multi-scale learning module(ML)module embedded between the upsampling and downsampling of the EAST model.The ML module adopts a split-transform-merge-residual architecture to extract multi-scale information from each layer.Secondly,in order to solve the problem that some texts are missed due to insufficient lighting or insufficient focus,this paper introduces a dual attention module to the EAST decoder to obtain better feature representations,so that these texts can be detected.Finally,this paper has made some improvements to the existing geometric loss function in scene text detection to improve the model’s ability to detect scene text areas.Extensive experiments prove that the proposed method is more robust than several other benchmarks in the Mongolian dataset. |
