Place: Large Lecture Room
Affiliation: Kyushu University, Japan.
Now OCR is extending its target from scanned documents to scene images. Please look around and find texts around you — there may be various texts printed in various fonts. Thinking about scene texts and their component characters will provide us many interesting problems, which never happened on the conventional OCR for scanned documents.
First, how about “non-verbal” information given by a specific font? For example, please recall the logo of a famous gorgeous brand. Is the font used in the log the same as that in a toyshop? The answer is, maybe, “no”. This suggests that each font has its own “atmosphere”, which is typical “non-verbal” information.
Second, how about “visual perception of scene texts?” Most scene texts are put in the scene for providing some message to people. They, therefore, should be “eye-catching”, or, visually salient. Recently, this visual saliency is a hot topic of computer vision and thus it is not bad to examine that scene texts are really salient or not.
Third, how about “the context of scene texts?” Scene texts are often related to their scene context. For example, on car roads, scene texts are often related to navigation information for drivers (i.e., traffic sign). Another type of scene context information is, for example, that if scene context is “sky”, there might be no text around it. This means the scene context gives a prior probability for scene character detection. In this talk, I will introduce those points along with my other research activities.