Research On Chinese Text Detection In Natural Scene Based On Deep Learning

Text detection in natural scenes is a technology to locate the bounding box of words or text lines in the image.The society's ever-increasing demand for artificial intelligence technology promotes the rapid development of artificial intelligence.As part of artificial intelligence,text detection has gradually become an indispensable technology in this process of development,and therefore get broad development prospects.At present,deep learning has become an important method of text detection.Multiple text detection algorithms based on deep learning have achieved good results and performance in practical applications.However,there are less research on Chinese text detection in natural scenes,and the existing text detection methods are fail to meet the needs of users.Thus this paper mainly does research on Chinese text detection based on deep learning.In order to improve the detection accuracy of text detection networks,three effective text detection methods are proposed:(1)A text detection method based on deep neural network multi-level feature fusion(abbreviation: Fusion CTPN)is proposed.Feature maps generated by multiple levels are added to the training process of CTPN.Firstly,multi-level feature map is generated by feature extraction network.Then,the large-scale feature maps generated at the lower level are combined with the small-scale feature maps generated at the higher level to get higher-quality feature maps.Thereby,the expression ability of the feature map has been improved.(2)A text detection method based on deep neural network multi-scale prediction(abbreviation: Multiscale EAST)is proposed.By constructing three output layers in the EAST network,the smallscale feature map is also added to the final prediction result.Then,the network learns more about the characteristics of small-scale text.The final multiscale prediction results and the corresponding real detection results act on loss function to realize multi-scale prediction.(3)A text detection method based on deep neural network loss function fusion(abbreviation: Balanced EAST)is proposed.The loss function Balanced loss is combined with the loss function of the network EAST by using weighted fusion to jointly constrain the optimization direction of the model.Thereby,the model's ability to learn complex features is improved.Similarly,the same loss function fusion method is applied to the text detection method proposed in the second point.The methods presented in this paper have been tested on both the Chinese dataset and the English dataset and are compared with some existing algorithms by using a unified evaluation standard.The final experimental results show that that the proposed method not only in English dataset,but also in Chinese dataset,which effectively improves the accuracy of text detection.