Text Detection In Natural Scene Images Based On Deep Learning

In recent years,with the rapid development of the artificial intelligence society,how to locate and identify texts in natural scenes is increasingly becoming a research hotspot in the field of computer vision and deep learning.Text localization algorithms mostly rely on artificially designed features,which leads to complex calculation and low efficiency.Some text location algorithms based on deep learning mostly target horizontal text detection problems,leading to poor positioning.In addition,the traditional Convolutional Recurrent Neural Network(CRNN)has insufficient ability to recognize the text image when the background and the text are similar in gray channel,and it also has poor detection effect on long sequence image texts with complex background noise between character.In view of the above problems,this paper will focus on the study of text detection based on deep learning in natural scenes image.Firstly,this paper has proposed a text-localized full convolutional network algorithm based on Convolutional Neural Network(CNN).The algorithm mainly includes: 1)Proposing a text-localized full convolutional network based on CNN in deep learning,the text features in images are automatically extracted to avoid the use of artificially designed feature defects;2)The sine and cosine values converted from the angles were added to the training of the text-localized full convolutional network,enabling the network to extract angle feature information of tilt texts in natural scenes image.3)Post-processing the feature information output from the text-localized full convolutional network algorithm,and predicting the coordinate positioning information of the text in natural scene image.The experimental results show that the text-localized full convolutional network algorithm based on CNN improves the localization ability of image texts,especially oblique texts in natural scenes,avoids the problems caused by artificially designed features,and improves the algorithm's calculating speed.Next,this paper also has proposed a Chinese text recognition network based on CRNN and attention mechanism,redesigning the network structure and added attention mechanism: 1)In the input of the CRNN,one channel used for the original grayscale image was converted to the three channels used for RGB images,which makes the CRNN have the ability of extracting feature information from the color information of the scene image text,avoidingthe problem of recognition errors caused by similarity between text and background in grayscale images;2)The first level Long-Short Term Memory(LSTM)of Recurrent Neural Network(RNN)in CRNN was added the attention mechanism that allows the RNN to focus on the text feature information in the long sequence of feature vectors.Experimental results show that the algorithm can solve the problem of insufficient recognition ability in long sequence text images with complex background noise between the text image and the text image when the background and the text are similar in gray channel.Finally,we combine the two algorithms proposed in this paper,aiming at the problem of broadcast control of scenic video content in a certain operator's platform business,using C/S network architecture to design a video-specific sensitive content monitoring system based on deep learning.The system can automatically judge in a short time whether the text in the returned video from the remote network monitors to contain sensitive content and save the corresponding detection result.The corresponding processing result can be inspected on the client in real time.Experimental results show that the system can locate various textual information and can accurately recognize and judge textual content.It can be applied to multiple scenes such as square monitoring and scenic spot monitoring.