Scene Text Detection With Stroke Width Transform And Convolutional Neural Network

Text detection and recognition in natural scene images is one of the important research in computer vision.It supports the application of text information such as license plate reading,multi-language translation,digital map creation,content-based image retrieval.Recently,the proposed algorithms have achieved good performance and the researcher released a series of evaluation datasets.However,there are still many challenges in scene text detection and recognition due to some factors such as complex background and abundant text.In addition,the number of images in the existing datasets is in small scale,and the image content can not really reflect the actual world.So it is necessary to build a more challenging dataset.The thesis mainly focuses on the text detection and recognition in natural scene images.And,the main works and novelties are as follows.(1)A dataset for scene text detection and recognition is presented,namely STV2k dataset(Street View 2,000).The number of images in the dataset collected from real scenes are large.The images are well-annotated,challenging and rich of contrast,orientation,physical appearances,etc.Compare with most of the existing benchmarks,our dataset is the better according to the various attributes of dataset(such as the number of images,character categories).Two outstanding methods are evaluated on our dataset.The experimental results are imperfect.When the dataset is published,the relevant researchers of the Microsoft Research Asia,Institute of Automation Chinese Academy of Sciences and Xiamen University request for academic research.(2)A scene text detection method with Stroke Width Transform and Convolu-tional Neural Network is proposed.The low-level features obtained by the Stroke Width Transform are combined with the high-level features based on the Convolu-tional Neural Network in our method.We evaluated the proposed method on three benchmarks which are the ICDAR 2003 dataset,the multilingual dataset and our STV2k dataset.Experimental results show that our detection method yields good performance on the benchmark datasets and its running speed is fast.