|
Text
recognition in digital video presents several new challenges over
traditional OCR of scanned documents. First, the text resolution
is often so low that commercial OCR software can not recognize
it reliably. Second, text is often embedded on complex background
so text separation from background is a difficult task. In this
talk we propose a super-resolution based method to increase the
textual image resolution and smooth the background. We make use
of the fact that the same text string usually exists in the consecutive
multiple frames and register them in the subpixel accuracy. The
registered text blocks are fused to achieve a new text block with
high resolution and clean background. We will describe the implementation
details and discuss the conditions our scheme supposes. Experiments
on several video sequences show that our enhancement scheme can
improve OCR accuracy considerably.
|
