Multi-scene Text Detection And Recognition

In recent years,with the progress of science and technology and people's improved living standards,the demand for image content understanding has increased.The application of text detection and recognition is not confined to only a single scene,but a mixture of multiple scenes.This leads to the requirement that the task of text detection and recognition in multi-scene should deal with complex illumination conditions,variable text fonts and angles,which greatly challenges the performance of the algorithm.At the same time,academia and industry also hold many text detection and recognition challenges in different scenarios,which also reflects the popularity in research on multi-scene text detection and recognition.There are three main works:(1)Based on the deformable convolution networks,it is improved and applied to the horizontal text detection task.The results in natural scenes and bills scenes show its effectiveness and it won the second place in ICDAR2017 COCO-Text text detection contest(2)From the perspective of instance segmentation,a multi-oriented text detection framework called FTSN is proposed,which combines low-level detail information and high-level semantic information to obtain fusion feature maps,and the post-processing method of non-maximum suppression is improved,so that it achieves the leading performance in multi-oriented text detection datasets such as ICDAR2015 and MSRA-TD500 which contain different scenarios,and it can be naturally extended to the task of curve text detection.(3)Based on CRNN,which is a popular text-line recognition method,DSAN is proposed to deal with text recognition.Spatial attention module and deeply supervised module are introduced to enable feature maps before sequence processing to focus on activating the semantic information part and suppressing the redundant and cluttered information part.This method achieves the leading results in many text recognition datasets in different scenarios such as ICDAR2013,SVT and IIIT5 K.The method proposed in this paper for multi-scene text detection and recognition tasks is validated in multiple datasets or contests,and reach the state of the art.