Research On Scene Text Detection Method Based On Deep Learning

Nowadays,people are in the information age,researchers are more and more keen to get the required information from the image,and the text on the image can help us better understand the image information.Therefore,extracting text information from images has become a popular research task in the scientific research field in recent years.However,images taken under natural scenes are very complex and changeable,which brings great challenges to the detection task.Traditional machine learning methods can no longer meet the needs of tasks.Thanks to the promotion and application of deep learning technology,the scene text detection task has brought new opportunities and possibilities.A large number of researchers have applied deep learning technology to the detection task and proposed a series of classic algorithms.Based on this,this paper mainly studies the scene text detection method based on deep learning,and proposes three different scene text detection methods,namely:(1)A Hierarchical Scene Text Detector concerning Hard Examples.In the past,many deep learning-based methods used to extract text features after multi-level extraction,and then predict and classify them on the last layer of feature maps,which would lose a lot of feature information in the middle layer.At the same time,the network training does not pay attention to the distinction between difficult samples and simple samples,resulting in insufficient training of difficult samples and low detection accuracy.To solve these problems,this paper proposes hierarchical scene text detector concerning hard examples,namely the HST-DHE method.This method is based on a fully convolutional neural network,predicts the text area pixel by pixel,eliminates the pre-defined anchor box,and make use of the feature pyramid for hierarchical prediction.At the same time,it incorporates the idea of hard example mining to redesign the loss function of focusing on difficult examples,so that the network can pay more attention to hard examples and further improve the accuracy of text detection.(2)Curve Text Detection based on Generative Adversarial Networks and Pixel Fluctuations.As segmentation-based methods often fail to produce satisfactory segmentation results,a scene text detection method based on generative adversarial network and pixel fluctuation is proposed,namely the GAPF method.This method introduces the generative adversarial network into the text detection field,and uses the generative adversarial network as the main framework to generate accurate text segmentation results.At the same time,the concept of pixel fluctuation is introduced,and the pixel fluctuation information of the image is input into the generator network as a condition to enhance the translation and rotation are invariable.Finally,a new postprocessing algorithm is designed to generate the boundary of the text from the segmentation result.(3)Scene Text Detection based on Weakly Supervised and Saliency Map.Aiming at the problem that the fully-supervised method is very time-consuming and costly to label data,a weakly supervised scene text detection method based on saliency map is proposed,that is,the WSSM method.This method is inspired by saliency target detection.It uses a segmentation network to generate a category saliency map of the text area on the scene image,and then on the basis of the category saliency map,the text area is labeled with a bounding box,and finally inputs the image with the salient map into the network for training,so as to replace manual labeling.