Research On Text Detection Technology Based On Improved Convolution Neural Network

Text detection is a technology that locates text region in the image,which is widely used in intelligent tourism,driverless and other tasks,and plays an irreplaceable role in all aspects of the intelligent society.The traditional text detection technology has many defects such as low detection accuracy,narrow applications,weak robustness,and reliance on artificial manual guidance.The deep neural network can obtain a detection model with strong generalization ability and robustness through the training of a large amount of data.In the field of text detection,the introduction of dilated convolution in the convolutional neural network as the backbone network can effectively increase the receptive field,and can further enhance the feature extraction.Therefore,the improvement of convolutional neural network technology has gradually become one of the important research directions in the field of text detection.In the natural scene text detection,especially the detection of Chinese artistic characters in natural scene is a difficult point in the field of text detection research.For example,Chinese cursive text has continuous strokes,and the characteristics of text in natural scenes are extremely irregular and the direction is variable.These text detection technologies have brought great challenges.At present,although text detection methods based on convolutional neural networks have achieved certain results,there are still some problems that have not been effectively solved in the task of detecting cursive characters and arbitrary shape text detection tasks in natural scene.In order to solve the above problems,this paper overcomes the current lack of cursive text detection dataset,constructs cursive detection dataset,and proposes a cursive text detection method based on self-built dataset.In view of the shortcomings of the current natural scene text detection,this paper also proposes an improved method of text detection of arbitrary shape.This paper highlights three points as follow:1.In view of the lack of cursive text detection and recognition dataset in the current text detection field,this paper designs and constructs a cursive dataset named Chinese Cursive Dataset(CCD).The dataset contains about 20,000 characters,and contains various features that may appear in real cursive scripts,such as continuous strokes,irregular fonts,and variable directions.A series of comparative experiments were carried out on the Cursive dataset to verify the research value and significance of the dataset proposed in this paper.2.In the detection of Chinese cursive characters,his paper proposes a text detection method based on self-built cursive dataset(CCD?cdm).This paper uses vgg-16 as the backbone network to extract features,and incorporates the attention mechanism block to further enhance the feature extraction capabilities of the backbone network.On the other hand,for the cursive characters with the characteristics of continuous strokes,the method of segment linking is adopted.Specifically,after a part of the character region was detected,more text regions are gradually detected than before,and a link is used to determine whether the detected region was a text region,and finally a single character region of cursive text is obtained.3.In the task of text detection of arbitrary shape in natural scene,we improve and propose an arbitrary shape text detection method in natural scene.For the problem that the residual network is not weighted from the perspective of feature channels,an attention mechanism is introduced to effectively mine the network to perform weighting operation on the feature channels to enhance the feature extraction capabilities.In addition,this paper introduces dilated convolution in the backbone network to improve the receptive field of the backbone network without changing the parameters of the network model,and incorporates Jaccard coefficient in the post-processing of the model,so as to adapt to the situation that the text only occupies a small part of the image area in the natural scene.In this case,our model improves the accuracy of text detection under arbitrary shape.This paper conduct a number of exact experiments in a great quantity of real datasets,and extensive experimental results demonstrates that our model achieves the best performance compared to some state-of-the-art text detection methods.