Uyghur Detection Based On Convolutional Neural Network

Uyghur is the main language used in Xinjiang.It is a popular language in ethnic minority areas.Most of the text appears in video images.It recognizes and understands these words to strengthen exchanges between the mainland and Xinjiang,maintain stability in Xinjiang,and promote economic development.Aspects are of great significance.In recent years,artificial intelligence and big data technologies have developed rapidly.Computers can obtain effective information by processing massive amounts of voice,text,video and images.This provides a new direction for accelerating the development of Xinjiang.In order to achieve the purpose of identifying and understanding Uighur,it is first necessary to detect Uighur in the picture and segment the area,ie target detection and semantic segmentation,to further understand the text.The text in the image is divided into two forms.One is that the text is embedded later.This type of text has a large feature difference from the background,and the detection difficulty is small.One is that the text exists in the natural scene,and the text is subject to The influence of natural environmental factors such as color size,shooting angle and illumination noise is difficult to detect.In this paper,a Uyghur text in natural scene images is proposed to extract and learn target features using convolutional neural networks to identify words.The convolutional neural network has powerful feature learning ability,is sensitive to target features,and has strong anti-interference ability,and is suitable for dealing with various interferences under natural conditions.In view of Uighur's letter irregularity and unique writing style,this paper explores Uighur detection and segmentation by using deep learning method.Firstly,based on convolutional neural network,a multi-layered feature network is constructed to obtain rich semantic information.The characteristics of the Uyghur language are determined by the regional recommendation network.The ROIAlign method is used to obtain the exact location.Then,the Uighur category,the border coordinates and the segmentation area are calculated by the target detection branch and the semantic segmentation branch respectively.Through experiments,the text network in the target detection branch and the semantic segmentation branch are better than the recent popular convolutional neural network,the accuracy and category accuracy reach 92.2% and 95.9%,and the intra-region consistency rate and pixel precision are 93.21% and 96.68%.At the same time,the method and several classical semantic segmentation algorithms are compared in the Uighur dataset,which is 12.9 and 13.6 percent higher than the best performing algorithm PSPNet,which verifies the good performance of the network and algorithm.In addition,since the existing international public dataset does not have a Uighur image collection,this article has done a lot of work on collecting,photographing and producing Uyghur natural scene photo collections.