Research On Local Feature Based Scene Text Analysis Methods

Text in scene images contains a wealth of valuable information that are key clues for intelligent control and information retrieval system. The number of video and image data becomes larger and larger along with the popularization of image acquisition equipment such as smart phone. There are urgent need for automatic approaches to analysis,recognize and extract scene text. In recent years, there has been a dramatic proliferation of research concerned with how to extract text information from the scene images,especially text with complex background. However, the division between the industrial requirements and the present technologies remains.Local feature based methods could overcome the above disadvantage to a large extent and because that local features are shift invariant, scale invariant, rotation invariant and affine invariant. Therefore, it is necessary to conduct a systematic research about the application of local feature in scene text analysis. Focusing on improving the performance of scene text analysis algorithm by local feature, the research work of this dissertation investigates the scene text analysis methods based on local features under different conditions.A video text detection approach based on open-door enrolling strategy is proposed.The proposed approach improves the recall rate by perform local feature based text detection in multiple scales under the condition that the video frame contains text in different scales. On this basis, a text region verification method based on stroke width feature is proposed. This method improves the accuracy of text region verification by analysis the distribution of stroke width in candidate regions. Moreover, a semi-supervised multiinstance learning algorithm is proposed to solve the problem that the polar parameter is unknown during the process of stroke width extraction. The experimental results proof that the proposed method is efficient with high recall rate and achieve high accuracy.The recognition of character in scene images is difficult under large category number and significant inter-class variation. The performance of tradition OCR methods is often far from ideal. The research work of this dissertation investigates the recognition of scene characters and proposed a character recognition method based on ensemble learning and model compression. Ensemble learning could improve the generalization ability of the classifier. On the other hand, the ensemble classifiers are effective but often inefficient. A Model compression method based on boundary samples and local classifiers is proposed against this. The proposed method could reduce the number of pseudo sample and compress the ensemble classifier into equivalent classifiers that are more efficient. Moreover,a feature extraction method based on local feature and spatial histogram is proposed. The combination of these two methods can effectively increase the accuracy of scene character recognition.A word image based scene text analysis methods is proposed under the condition that some scene text that is difficult to be segmented into single characters. The proposed method code local features by utilizing balanced randomized forest and then calculate the feature vectors of word images based on bag-of-feature model. Thus the proposed method could generate effective lexicon without supervised information and overcome the problem that word image samples are difficult to collect. The experiment that cluster word images proof the proposed method is highly reliably.The results of text detection are not accurate enough in complex scene images. To against this, this dissertation proposed a scene text analysis method based on constellation character model. The proposed model describes the whole character using collection of local features. The appearance of local features and the relative position among features are modeled by probability model separately. Thus the probability that the target character presents could be calculated. The proposed method is more flexible compare to global feature based methods and is more effective in complex scene images.