Integrated Coarse-Fine Scheme for Text Information Extraction from Instructional Videos

In this paper, we propose an effective coarse-to-fine algorithm to detect and extract text in instructional videos. Firstly, in the coarse section, an edge-based algorithm is employed to detect all candidate regions of character edges. First in the detection step, an edge map is created using the canny edge detector. Then, morphological filtering is used, based on geometrical structure element, in order to connect the vertical edges and discard false alarms. A connected component analysis is performed to the filtered edge map in order to determine a bounding box for every candidate text area. Finally in the localization step, horizontal and vertical projections are calculated on the edge map of every box and a threshold is applied, refining the result and splitting text areas in text lines. Secondly, in the fine section, correct text regions are selected from candidate ones by support vector machine (SVM) model and texture features. Finally, we segment these regions and binarize them to be fed into the OCR engine to be recognized. Experimental results show that our algorithm achieves high performance and prove that our system is highly effective and efficient for text information extraction.