Research On Natural Scene Text Detection Technology Based On Deep Learning

Words are an important advancement in the history of human civilization and a sign that humanity has entered the age of civilization.They are widely present in natural scene images.People can understand the external information they need based on the text content in natural scene images,such as their geographic location and the direction of the vehicle.Compared with the cumbersome expressive ability of pictures,the expressive ability of words is more conceder and clear.At present,natural scene text detection technology based on deep learning has been widely used in all aspects of life,such as intelligent wearing devices that can take pictures and obtain information anytime,anywhere,mobile phone photo translation function,and unmanned driving.At present,due to the rapid development of deep learning technology and the improvement of technological level,text detection technology has also been greatly improved,and has achieved some results.Various excellent text detection algorithm models are constantly being proposed and improved,and the text detection effect is getting better and better.In addition,various text detection and recognition contests are held every year to attract countless outstanding scientific researchers to participate.In this thesis,in view of the problem of inconsistent size of slanted text in natural scene images and the problem of low accuracy of text detection in complex scenes,on the basis of an efficientand accurate scene text detector(EAST)algorithm,a one-stage text detection algorithm is proposed.The algorithm uses reinforcement learning to train a recurrent neural network controller,select the optimal full convolutional neural network structure,and extract the multi-scale features of the text.Then import it into the output module and use the generalized intersection over union algorithm to enhance the regression effect of the text bounding box.Finally,through the modification of the loss function to ensure positive and negative sample category balance.The algorithm achieves 0.82,0.88,and 0.85 in the ICDAR2015 data set for the recall rate,accuracy rate,and comprehensive index respectively,which can effectively detect slanted text in natural scene images.Finally,in order to solve the problem of curved text in natural scene images,this thesis designs a two-stage text detection algorithm.The algorithm makes full use of the shallow features extracted by the model by adding new paths to the feature pyramid network used for feature extraction.Then,the adaptive feature pooling algorithm is used to pool the features of various scales to extract richer features of the region of interest.Finally,by using the progressive scaling algorithm to effectively separate adjacent text,irregular text detection areas are generated.The experimental results show that the recall rate,accuracy rate and comprehensive index in the SCUT-CTW1500 data set reach 0.75,0.81 and 0.78 respectively,which can effectively detect the curved text in the natural scene image.