Text Detection In Natural Scenes

Text information extraction in natural scene in the field of computer vision is a very important and challenging problem, it has so broad application prospects in the field of image retrieval, assisting blind translation system and intelligent transportation that has received a significant amount of attention from companies and researchers.Text information extraction in natural scene include the detection, segmentation of text and recognition of character, text detection is the first step of text information extraction, the accuracy of text detection has important significance for the subsequent segmentation and recognition. This paper will focus on text detection. Because deep learning has been successfully applied in the field of image classification, therefore this paper is under the framework of deep learning, put-forwarding supervised text detection method based on convolutional neural network and unsupervised text detection based on Autoencoder neural network, this improve the accuracy of text detection.The main work in this paper are as follows:(1) This paper design semiautomatic text annotation tool which can easily generate training and testing data sets, we use the tool to create a text detection bench mark data sets. Because of the deep learning framework using multilayer network model, which requires a lot of input as the training set, the existing text detection training set cannot meet the requirements and usage is not convenient, therefore, this paper design a semi-automatic text annotation tools, can quickly create large training data set. In this paper, we collect a part of ICDAR2003and SVT data sets, including4127images, to generate60000positive samples and71733negative samples, which are used for training,10000samples are used for testing, so it can greatly shorten the period of the experiment, providing the basic data for the performance evaluation of different algorithms.(2) We propose a kind of supervised algorithm of learning feature based on convolutional neural networks to detect text in natural scenes. CNN(convolutional neural network) is a kind of multilayers perceptron, the network structure can keep high degree invariance with translation, zoom, tilt or other forms deformation, so as to improve the accuracy of text detection. It achieves the classification accuracy of93.56%on the test data set in the experiments of this paper.(3) We propose a method of unsupervised algorithm of learning feature based on Sparse Autoencoder to detect text in natural scenes. The method is a kind of feature learning to minimize the cost of the reconstruction error, adding the data sparse constraints on Autoencoder. It can obtain more effective feature representation. This method achieves92.85%classification accuracy on the test data in this paper.(4) We compared supervised feature learning algorithm with unsupervised feature learning algorithm through analyzing the experimental results, we can draw the following conclusion:The classification result of text detection of supervised feature learning based on CNN, slightly better than supervised feature learning for text detection based on SAE(Sparse Autoencoder), it may be due to SAE for adding sparsity constraint, but SAE is quicker than CNN on the speed, on the other hand, because SAE is unlabeled data, therefore, it has much room for improvement.