Research On Natural Scene Text Detection Algorithm Based On Deep Learning

With the advent of the information age,mining more useful information from big data has become a value orientation.The text is an important channel to obtain information.In the natural environment,text information is ubiquitous.With the continuous development of computer vision technology,the conversion of text information in natural scenes into digital information will improve the machine's ability to understand the scene.Important role.The rapid development of neural networks makes detection algorithms based on deep learning the mainstream method.Through studying the mainstream algorithms in the field of text detection,this paper sorts out the processes and methods of various text detection algorithms,mainly including text detection algorithms based on target detection and text detection algorithms based on instance segmentation.The advantages of the two types of algorithms are analyzed.After the shortcomings,based on the differentiable binary text detection algorithm,a text detection algorithm based on text feature enhancement and text feature adaptive fusion is proposed.The main work of the algorithm is as follows:(1)In order to solve the problem of the loss of semantic information in the natural scene text detection algorithm based on segmentation,which leads to the missed detection of large-scale text,a feature enhancement network is proposed.By injecting more spatial context information into the original branch of the feature pyramid,the information in the top-level feature map is supplemented and enhanced,and the information loss of the top-level feature in the downward propagation process is reduced,so that the feature pyramid can obtain more semantic information.Better guidance for segmentation improves the segmentation accuracy of the network.(2)In order to improve the conflict of different levels of text features in fusion,a text feature adaptive fusion network is proposed.Texts of different scales and sizes will be mapped to different levels of features after passing through the convolutional network,which may cause the text samples to be marked as positive samples in the feature maps of a certain level,but they are In the case of negative samples,directly fusing these features will get a sub-optimal result.The adaptive text feature fusion network obtains the spatial weight matrix of each feature map through learning,and then uses the weighted fusion method to generate the better feature map required by the segmentation network.Finally,this paper verifies the algorithm on the disclosed multi-directional text detection data set ICDAR2015 and the curved text detection data set Total-Text detection data set.The results show that on the Total-Text data set,the feature enhancement model and adaptive feature fusion network proposed in this paper increase the comprehensive indicators of the baseline by 1.36 percentage points and1.1 percentage points,respectively.On the two data sets,compared with the current mainstream algorithms,it also has a certain degree of competitiveness,especially the comprehensive detection index for curved text achieved 85.4%.This verifies the effectiveness of the optimization algorithm.