Computer Vision Research Group

Text Detection in Videos and Natural Scenes

In natural scene images, texts provide highly useful information and can provide a key for understanding the image content. Similarly, the rapid growth of video data creates a need for efficient content-based browsing and retrieving systems. Text in various forms is frequently embedded into images to provide important information about the scene like names of people, titles, as road-sign Identification, interpretation of documents such as geographic maps or engineering drawings, locations or date of an event in news video sequences and natural scene images, etc. Therefore, text should be detected for semantic understanding and image indexation.

We have performed the detection of text in a video frame or a natural scene image in few major steps such as pre-processing, features extraction and classification.

Pre-processing is applied to identify the candidate regions. Edge information is extracted using the optimal edge detector. This helps in finding connected components precisely by removing useless edge pixels. After finding connected components, character block filtering is applied to screen those connected components that do not contain texts.

Text extraction and verification is done by using features extraction. Currently, we are working with two sets of features; the first feature set uses the discrete wavelet transform (DWT) to extract the features of characters for texture analysis and description. The second feature description is a new one that is local energy based shape histogram 'LESH'. In order to provide training samples to the feature extraction step, we have collected a data set of training samples that contains 2000 text images and 650 non text images gathered from different images and video streams.

The classification is currently being done using nearest mean classifier; however, we are working with adaptive adaboost algorithm for classification purpose to enhance results.

Recent News
© 2010 Department of Electrical Engineering, CIIT Lahore  |  Terms and Conditions