Research On Distortion Text Detection Method Based On Convolutional Neural Network

With the development of AI,computer vision technology is also applied to all aspects of life.Scene text detection is a basic task in the field of computer vision,which is usually combined with text recognition,text extraction and other technologies,and applied in the fields of smart city,intelligent assisted driving,blind guidance for the visually impaired and so on.At present,the general detector is mainly used to detect document graphics,but the complex background environment and the distortion of scene text lines will lead to a large number of missed detection and false detection problems.In order to overcome the complex practical problems such as scene text distortion and improve the efficiency and quality of text detection.In this thesis,combined with the general detection model and convolutional neural network theory,the scene text detection method based on convolutional neural network is studied.The main work of this thesis is included as follows:(1)Comparing the existing network model,the high efficiency on feature extraction by convolution neural network,and the extracted features can better express semantic information.The convolution neural network is applied to the scene text detection network, and the extracted scene text features are more robust,and are applied to the next stage of scene distorted text detection.(2)Aiming at the problems of distorted text lines in natural scene and low detection quality caused by complex environment,this thesis improves the network based on the general detector model at present,and proposes a scene distorted text detection algorithm based on PSE network structure.The backbone network is replaced by a lightweight network to improve the speed and quality of model detection.When dealing with the conflict of overlapping pixel values of adjacent text boundaries,K-nearest is used to replace the queue mechanism to make the training speed more faster.(3)In order to solve the problem of imbalance between efficiency and quality of scene text detection by conventional method,a cascade scene text detection method is obtained by improving the PA network.Firstly,the text features of scene are extracted by using Mobilenetv2.Then,build a cascade network which makes full use of the context semantic information,and enhance the feature information on the basis of the initial feature extraction of mobilenetv2.Finally,the high-level network features which are easy to classify and the low-level network features which are easy to locate are fused to make the feature extracted by cascade network more robust and easier to locate and classify the distorted text area.Through the experimental verification on the open data set ctw1500 and icdar2015,the performance of this method in scene distortion text detection has been improved.The algorithm based on improved PSE network structure not only has better accuracy,but also has better detection effect when the shape of scene text line is distorted.For the method based on PA network,this method not only keep a good accuracy,but also reduces the network model parameters,so as to improve the detection speed and achieve real-time detection.