Automatic caption text detection and processing for digital images

Abstract

A texture-based text localization system proceeds directly in the compressed domain for DCT compressed JPEG images or MPEG videos. The DCT coefficient values in JPEG images and MPEG videos, which capture the directionality and periodicity of local image blocks, are used as texture feature measures to classify text areas. Each unit block in the compressed images is classified as either text or nontext. In addition, post-processing in both the compressed domain and the reconstructed candidate text areas can be used to refine the results. For video frames that contain text, the displacement of text between two consecutive frames is estimated which gives the velocity of the moving text. This temporal displacement information is also used to further refine the localization results. The text is then processed to provide content or speech output.

Dependencies are inferred from the claim text.