Multi-Oriented Video text detection and extraction using DCT feature extraction and Projection based rotation calculation

With the popularity of Smart phones, Handy cam, CCTV etc with sophisticated software´s and good resolution cameras Scene Video text detection has gained considerable interest among researchers. Scene Video text is challenging because it has to deal with texts of arbitrary orientation, variable size and uneven illumination. While scene text detection has generally been treated with a notion of horizontal alignment, our proposed method aims to detect Multi-Oriented Scene Video text by using the knowledge of block wise AC coefficients in Discrete Cosine Transform and verifying the result using Stroke Width calculation. The orientation angle is found by Rotation calculation by maximum variance estimation method to localize Text region connected components. False positives are eliminated by applying various heuristics. Adopting this approach on various standard and own databases and comparing with some established state-of-the-art method proves its effectiveness and robustness to various Text anomalies like sizes, fonts and orientation.