Research On Text Detection Algorithm Of Natural Scene Based On Deep Neural Network

Text detection in natural scenes is an important part of computer vision,and it also brings us convenience,such as identification of ID photos,unmanned driving technology,intelligent robot technology,and so on.For now,printed text recognition is quite mature and widely used in our lives.With the widespread use of media technology nowadays,the background of images has become more complex.Due to the complexity of the natural scene and the characteristics of the text itself,the use of deep learning can more accurately and quickly detect the text lines in the image,so as to achieve the purpose of text detection.How to design an algorithm that can cope with the challenges of various environments when detecting text and make the detection effect better.This paper is mainly based on the deep learning network framework,and aims to design an efficient and fast text detection algorithm.The main research contents of this paper are as follows:1.In the improvement process of EAST,based on the hole convolution algorithm,this article adds the ASPP structure to the EAST network to increase the detection field of view,to achieve long text,multi-angle text detection,and improve the detection effect.Taking into account the role of context information in feature fusion,after the feature fusion layer,the BLSTM algorithm is used to output sequence features,and the correlation between context information is used to estimate the text area and improve the text detection effect.Then,for the problem that the positive and negative weights of EAST samples are unreasonable,the balanced cross-entropy loss function is replaced by the addition of the two functions of Dice Loss and Focal Loss.Through experiments on the ICDAR2015 data set,the results show that the improved strategy is 6 percentage points higher than the standard EAST in accuracy,5.2percentage points higher in recall rate,and 5.7 percentage points higher in F value.Compared with other experimental algorithms,the overall performance has been improved.2.This article uses the lightweight neural network Mobile Net V2 as the backbone network.Large-scale network feature extraction can get rich features.However,because the model may be too large in the extraction process,this chapter uses light-weight networks to replace large-scale networks to solve the problem of excessive parameters and reduce application delay.3.Directly replace the large network in the EAST network feature extraction layer with a lightweight network,and its detection effect will be greatly discounted.Therefore,this article adds a hollow convolution module to the feature fusion layer,and merges the features of the two parts.,Use the channel attention mechanism to merge and screen the features to improve the use efficiency of features in the network.Finally,this chapter uses the public ICDAR2015 data set,and through experimental comparisons,the effectiveness of the improvement strategy in this chapter is verified.Finally,the overall performance of the EAST model introduced above is compared,which further verifies the rationality and effectiveness of the improved network in this paper.