Augmented edit distance based temporal contiguity analysis forimproved videotext recognition

摘要

Videotext refers to text superimposed on video frames and itenables automatic content annotation and indexing of large video andimage collections. Its importance is underscored by the fact that avideotext-based multimedia description scheme has recently been adoptedinto the MPEG-7 standard. A study of published work in the area ofautomatic videotext extraction and recognition reveals that, despiterecent interest, a reliable general purpose video character recognition(VCR) system is yet to be developed. In our development of a VCR systemdesigned specifically to handle the low resolution output from videotextextractors, we observed that raw VCR accuracies obtained using variousclassifiers including kernel space methods such as SVM, are inadequatefor accurate video annotation. We propose an intelligent postprocessingmechanism that is supported by general data characteristics of thisdomain for VCR performance improvement. We describe temporal contiguityanalysis, which works independently of the raw character recognitiontechnique and works well even for moving videotext. This novel mechanismcan be easily implemented in conjunction with VCR algorithms beingdeveloped elsewhere to offer the same performance gains. Experimentalresults on various video streams show notable improvements inrecognition rates with our system incorporating a SVM-based recognitionengine and temporal contiguity analysis