Video text recognition using sequential Monte Carlo and error voting methods

摘要

This paper addresses the issue of segmentation and recognition of text embedded in video sequences from their associated text image sequence extracted by a text detection module. To this end, we propose a probabilistic algorithm based on Bayesian adaptive thresholding and Monte-Carlo sampling. The algorithm approximates the posterior distribution of segmentation thresholds of text pixels in an image by a set of weighted samples. The set of samples is initialized by applying a classical segmentation algorithm on the first video frame and further refined by random sampling under a temporal Bayesian framework. One important contribution of the paper is to show that, thanks to the proposed methodology, the likelihood of a segmentation parameter sample can be estimated not using a classification criterion or a visual quality criterion based on the produced segmentation map, but directly from the induced text recognition result, which is directly relevant to our task. Furthermore, as a second contribution of the paper, we propose to align text recognition results from high confidence samples gathered over time, to composite a final result using error voting technique (ROVER) at the character level. Experiments are conducted on a two hour video database. Character recognition rates higher than 93%, and word error rates higher than 90% are achieved, which are 4% and 3% more than state-of-the-art methods applied to the same database.