HMM-based approach for text region detection in coded video bitstreams

This paper proposes a novel text region detection technique based on a hidden Markov model (HMM) for MPEG-encoded bitstreams. Although enormous number of techniques have been proposed for a detection or a localization of text regions, an HMM-based approach has not been proposed as far as the authors know. First, two kind of feature value, i.e., prediction-mode-based and bit-amount-based ones, are extracted as temporal sequences from a target MPEG bitstream, which are then combined together to be fed into an HMM. names containing text regions in the bitstream can be detected directly from the state transition sequence by the HMM. Experimental results have demonstrated that the proposed technique achieves a precision of 92% and a recall of 75%.