Research On Natural Scene Text Detection Technology Based On Convolutional Neural Network

Text detection of natural scenes is the realization of text detection and localization from natural scenes,and thus services and applications have an increasingly important role in all aspects of society.There have been some studies using convolutional neural networks to detect text,but there are still many problems that remain unresolved.This paper proposes a new method for the detection of long text and stone text in natural scenes.The main research contents of this paper are as follows:1.The long text detection for natural scenes because the extreme aspect ratio of text lines makes most networks unable to detect completely.In this paper,the feature excitation module is used to extract the channel information in the network model to suppress the background features and enhance the foreground features.A new elliptic geometric regression method is proposed,which can accurately complete the long text link without adding extra steps.And angular rotation to improve the accuracy of long text detection.2.The problem of the stone tablet text in the natural scene cannot be accurately detected because the "big" of the text exceeds the maximum setting of the text on most networks.The deep convolutional neural network is used as the feature extraction tool,and the extracted features are subjected to image information entropy calculation,and the value of the maximum entropy is taken as a part of the text candidate region.On the other hand,the feature weakening algorithm is used to eliminate the effects unrelated to the color features,and to input a part of the maximum stable extremum region algorithm or the text candidate region.The final result of the two is combined to obtain a text area feature map.Combine migration learning to train a classification network,and input candidate regions into the network to get the final text region.In the process of this task research,a stone dataset was created,two different tags were designed to support the research task,and compared with other most advanced ones on this dataset,the challenge of this dataset was verified.The research results obtained in this paper:1.The improved network model of the long text model in the natural scene has strong migration ability and can be used for research in other fields.The proposed model is compared with several state-of-the-art models.The best performance results were achieved in both accuracy and recall.2.The stone text detection model in the natural scene combines deep learning and traditional methods,and achieves the best results in comparison with other methods in small dataset.The proposed dataset is challenging enough to have room for expansion.