Deep Graph Learning Based Approach for Identification of Text in Scene Video
Mortha Manasa Devi · Journal of Emerging Technologies and Innovative Research · 2019
Content based analysis, retrieval, searching of scene video has become a key area under computer vision. Apart from indexing and retrieval of videos, demands for video analysis to monitor illegal videos have revolutionized the text detection problem. Because of complex background, low contrast, illuminated, variable font sizes, traditional approach of video based Optical Character Recognition (OCR) system performs satisfactory to detect the text from video. Later, two state-of the-art methods like SIFT and MSER outperformed to detect the text in video but both of these methods fails to detect with complex background. The proposed architecture utilizes the deep graph learning model to detect and identify the scene text from video in two stages. First, regions of similar nature are extracted from the frames by applying undirected graphs. Second, the extracted regions are fed to the learning model to obtain the features which are convolved with internal layers to find the probability of existence of text by calculating the gradients and gray level contrast between text and background. Compared to the conventional detection methods like SIFT and MSER, the detection rate based on deep graph learning can reach 90%. Experimental results show that proposed method is effective compared to two state-of-the-art methods SIFT and MSER.