A Study Of Text Detection In Natural Scenarios Based On Deep Learning

As technology products such as mobile phones and digital cameras are widely used in people's lives,image information has become one of the main ways to obtain information in our daily lives,and the text content in images has become a key source for people to perceive image information.More and more researchers are focusing on content-based image analysis.Effectively obtaining text information from natural images plays an important role in the classification,understanding and storage of image content.The text detection technology of natural images has important research significance and value.In recent years,artificial intelligence technology has achieved long-term development,and its application fields have become very extensive,especially in the field of imagebased target detection.The target detection technology with deep learning as the core extracts the feature maps in the image through a convolutional neural network(Convolutional Neural Networks,CNN),and uses a post-processing module to classify the image.There is no need to set parameters manually.The corresponding model can be obtained through the data set.Very efficient detection target.This thesis studies the text detection method of natural scene image based on deep learning.The main research work is as follows:1.This thesis first studies the mainstream detection algorithms and technologies in the field of text detection,and analyzes some classic domestic and foreign text detection algorithms in detail,including text detection algorithms based on traditional algorithms and text detection algorithms based on deep learning.Pros and cons.On the basis of this,it is determined that the EAST algorithm is the main object to deeply study the optimization method of natural scene text detection.2.Considering that the text context has a strong correlation,the text proposes an algorithm that serializes the fusion features of the EAST algorithm at the fusion feature stage to obtain sequence features.The algorithm takes the sequence features as input and inputs it into the BLSTM network.Form a loop before and after the text above and below the information to improve the problems of the network parameter convergence curve of the EAST algorithm that fluctuates greatly and affects the convergence performance.In addition,when the algorithm performs long text detection,when the adjacent area of the detection area is a character area,the probability that the detection area is a text line is increased according to the relevance of the text,thereby improving the detection ability of the algorithm for long text.Through comparative experiments,the algorithm in this chapter has improved accuracy by 1.50%,recall rate by 3.57%,and F value by 3.44% compared to the traditional EAST algorithm.3.In order to more accurately detect long text lines and merge text boxes to avoid overlapping detection boxes,this paper improves the EAST algorithm from another angle and proposes an improved network structure based on segmentation algorithm.First,the Res Net feature extraction network is used to replace the PVANet network in the original EAST algorithm,which not only simplifies the structure,but also can extract more complete feature information.Secondly,the Atrous Spatial Pyramid Pooling(ASPP)module is improved,and the codec-decoder structure combining the FPN and the improved ASPP module is constructed to improve the receptive field of the network to achieve the effect of accurate segmentation.Finally,the detection frame merging algorithm has been improved,so that the improved algorithm can more accurately merge text boxes,and at the same time have a certain improvement on the detection of long text.Through comparative experiments,the algorithm in this chapter has an improved effect on the detection of long text and the accuracy of text box positioning.