Research On Text Detection Algorithm In Natural Scene

Text information is an important clue in image.Detection of text in image is one of the important tasks in intelligent information processing.Traditional text detection tasks are usually used to extract text from document images.Because the background of the document is generally relatively clean and does not have complex texture information,text information extraction is often relatively simple.Under natural scene,however,the difficulty of the text extraction often far more than this,one reason is that the background is very rich,such as streets,vegetation,building,etc.,and text representation is various,art words,deformation,multi-language mixed,the imperfect words,light cover,etc.Moreover,the quality of images in natural scenes is often inferior to that of images in traditional documents.It is precisely for these reasons that the effect of text detection in natural scenes is not satisfactory.The research contents of this paper mainly include the following two points:1.According to the accuracy of the text detection problem,the paper studies the tilted a CTPN based text detection algorithm,the paper improved the text box of CTPN network fitting method and the text box grouping strategy,and behind the CTPN network series CRNN text recognition network fault detection results,so it can accurately detect tilted text in natural scene.The improved CTPN algorithm can accurately frame out the inclined text region in the natural scene,and the test accuracy on MSRA-TD500 data set reaches 0.84.2.For the quickness of text detection problem,this paper studies a kind of based on the level of the YOLO-tiny text detection algorithm,the paper in general,on the basis of target detection framework YOLO-tiny,by improving the Anchor Box aspect ratio and change the size of convolution kernels,the migrated YOLO-tiny text detection task,to be able to quickly detect tilted text in natural scene.Improved YOLOv3 tiny more suitable for text detection task,the final in ICDAR2013 data set on the testing accuracy is 0.532,the proposed algorithm has no advantage compared with other algorithm accuracy,but it has a fast detection speed and detection on the CPU speed of 30 FPS,and the algorithm more lightweight and can be deployed to mobile devices,the paper will finally the algorithm run successfully transplanted into the Android mobile terminal equipment.The two algorithms studied in this paper have different emphases.The skew text detection algorithm based on CTPN is more accurate but slower.Although the horizontal text detection algorithm based on YOLO-Tiny is not high in accuracy,the detection speed is very fast.In reality,different solutions can be selected according to the application scenarios.