On The Research Of Text Detection In Natural Scene

As an important carrier of information, texts have a high degree of generality and a good description ability. Therefore, the natural scene text detection plays an important role in intelligent transportation, mobile seeing-eye and image retrieval, etc. Different from the texts in the printed documents, natural scene texts are usually embedded in the complex backgrounds. Some impact factors, such as text font, color, illumination variation, etc, make the natural scene text detection to be a challenging research work. For the research status and wide application prospect of this technology, the natural scene text detection method is researched in this paper, and the research achievements mainly include the followings:Firstly, as license plate is a special natural scene text region, a license plate detection method based on the gradient information and cascade detection is proposed for the traf-fic monitoring image. In this method, the gradient image is obtained to unify the text and background contrast of different types of license plates. In order to effectively and quickly detect the license plates, a cascade detection framework is designed in our work. The cas-caded detectors and the simple rectangle features are used to eliminate a large number of backgrounds firstly. And then the effective verification schemes are designed to verify the detection results obtained in the previous stage. Experiment results on our traffic monitoring image database show the effectiveness of the proposed method.Secondly, a license plate detection method based on the detection and tracking strat-egy is designed for the traffic monitoring video. By using the spatio-temporal information between adjacent frames, and by introducing the mechanism of the target tracking, the poor detection performance of the license plate detector will be made up for the low resolution. In addition, the tracking results and the detection results will be compared in each frame to gain new targets, and the results with higher confidence value will be regarded as the final results to improve the drift problem caused by the long tracking processing. Experiment results on our traffic monitoring video database verify the superiority of this method.Thirdly, an effective text detection framework is designed based on the connected com- ponents and the texture features for handling the general natural scene text detection prob-lems. In this framework, in order to improve the result of the stroke width transform, the multi-level edge detection method is proposed to improve the accuracy of edge detection. In order to retrieve the missing texts, the "key region" is built based on the detected texts, and the judgment rules are designed based on the detected texts and visual context. The missing texts will be regained if they meet the criteria. Moreover, we devise the sub-regional level verification strategy to deal with the length variation of the text candidate. The experiment results on the ICDAR2005, ICDAR2011databases verify the effectiveness of this method.Fourthly, if the natural scene text is analyzed in isolation, the wrong classification result may be more likely obtained because of some negative factors. For these, a natural scene text detection method based on the confidence mapping model is proposed. This method proposed here not only uses the text-likelihood of text candidate itself, but also embeds similarity information between the adjacent text candidates to enhance the robustness of the scene text classification performance. The experiment results on the ICDAR2005, ICDAR2011, ICDAR2013databases show that the satisfactory results have been achieved.Finally, a descriptive model based on the descriptive mid-level patches is designed to effectively verify the text line candidate. We regard each text candidate in the text line can-didate as the mid-level patches, and mining the local and the overall description information of the text line candidate. In the component level evaluation, we design a bootstrapping recognition model of the text candidate. Furthermore, the sub-regional identification strate-gy is developed to decrease the severe texture distortion caused by scale normalization in the region level evaluation. The experiment results on the ICDAR2005, ICDAR2011, ICDAR2013databases show that the proposed method can effectively classify text line candidates, and it can also achieve competitive text detection results.In addition to the achievements presented in this paper, we discuss the imperfect aspects in our study and introduce the task in our further work.