Research On Text Detection In Natural Scenes Based On Deep Learning

With the rapid development of the electronic information industry,people's desire to understand this world through various electronic devices becomes more and more urgent.Text detection in natural scenes plays a crucial role in text recognition,and text recognition provides advanced language information for understanding the current scene.Different from the traditional Optical Character Recongnition,which analyzes and recognizes the image files of text data,the text detection in natural scenes is aimed at different types of text images,which undoubtedly increases the difficulty of detection.In recent years,the rapid development of deep learning has provided new methods for image detection.This paper uses the ICDAR text data set to study the following two aspects of text detection in natural scenes based on deep learning:1.Text detection method in natural scene based on instance segmentation neural network.The semantic segmentation neural network classifies each pixel in the image.The instance segmentation neural network classifies the number of each detection target again based on the semantic segmentation neural network pixel classification.The examples in this article segment the neural network,use the convolutional neural network to extract the features of the area to be detected,and then fuse the low-level spatial location features and high-level semantic information features through the full convolution network to classify each pixel in the image at the same time,the link between the current pixel and the surrounding 8 pixels is predicted.When the pixel type is text and its links with the surrounding 8 pixels are all positive,the area is connected.This method effectively solves the problem of the complex and changeable background of the target area in natural scene text detection,on the ICDAR2015 data set,comparing with the Seglink algorithm improves the F value by 8.7%.2.Text detection method in natural scene based on deformable convolution and attention mechanism.Different from general object detection,text objects in natural scenes have more irregular shapes.Aiming at this problem,this paper proposes a text detection method based on deformable convolution.The deformable convolution expands the spatial sampling range of the convolution by adding an offset,and solves the defect that conventional convolution can only be sampled at a fixed position.It does not require additional supervision items and can be trained directly on the target task.Aiming at the problem that the network is too deep for feature extraction,based on the natural scene text detection of the attention mechanism,this paper proposes to add the attention mechanism in the process of constructing the basic feature extraction network.The attention mechanism simulates the way humans process visual information.For a scene that only cares about the area to be detected,the attention mechanism can be used to obtain weighted feature values,which improves the performance of the text detection method in the natural scene of this paper.Compared with the simple instance segmentation network,the network with deformable convolution and attention mechanism improved the F value by 1.1%.