Uyghur Text Detection In Natural Scene Images Based On Deep Learning

As the most direct representation of human high-level semantic information,text plays an indispensable role in image understanding and has a wide range of practical applications.For example,there are wide application prospects in drone cruising visual obstacle assistance systems,automatic driving and network content monitoring and purging.Therefore,scene text recognition has attracted much attention.However,the mature optical character recognition technology is not suitable for the recognition of natural scenes due to the complex background of scene images,non-uniform care,low light contrast,text diversity,and image text perspective distortion caused by photographing.Scene,recognition of scene characters has become an important research direction in the field of computer vision.However,most of the existing methods are oriented to Chinese and English recognition,and there are few studies on Uighur recognition.The detection of text in natural scenes is an early stage of text recognition.Its purpose is to determine whether there are texts in images of different scenes(warning signs,street signs,etc.),and if so,locate the text.Due to the lack of open and labeled Uyghur sample sets in natural scenes,it directly affects the efficient and rapid development of scene Uyghur detection and recognition.In this paper,in order to solve this problem,the method of deep learning and image segmentation is used to generate artificial Uighur image sample data sets.The experimental results show that the adopted deep learning and image segmentation methods are effective,and the artificial Uighur sample data generated is very real and the words can be naturally incorporated into the natural scene pictures,and the nonmanual annotation of Uyghur language detection in natural scenes can be efficiently and quickly provided.At the same time,according to the unique characteristics of Uyghur characters,this paper improves the single-layer deep neural network structure for extracting the multi-level and multi-scale features of natural scene Uyghur.According to the characteristics of Uyghur text lines in natural scenes,a multi-scale specification,multiple aspect ratio default frames are designed to suit the needs of Uyghur language detection in natural scenes.An improved single deep neural network consisting of a Uyghur character extraction component and Multi-feature fusion text detection components which are trained to predict Uyghur text box position and text confidence in an end-to-end manner.The natural scene Uyghur text detection experiment shows that the improved single deep neural network method considers the impact of multiscale and multi-level registration on the detection accuracy.The accuracy and F-value of the algorithm are 0.7234 and 0.6115 respectively,which improves the accuracy of detection.