Camera Text Recognition based on Perspective Invariants

As camera resolution increases, high-speed non-contact text capture through a digital camera is opening up a new channel for document capture and understanding. Unfortunately, perspective and geometric distortions in camera image of documents make it hard to recognize the document content properly. In this paper, we propose a character recognition technique, which is capable of recognizing camera text lying over a planar or smoothly curved surface in perspective views. In our proposed method, a few perspective invariants including character ascender and descender, centroid intersection numbers, and water reservior are first detected. Camera texts are then recognized using a classification and regression tree (CART) structure. Experimental results show our method is fast and improves recognition performance greatly