A word shape coding method for camera-based document images

Abstract

This paper reports a word shape coding method to facilitate retrieval of camera-based document images without OCR. Due to perspective distortion, many reported word shape coding methods fail on camera-based images. In this paper, the problem is addressed by approximating the perspective transformation with an affine transformation, and employing an affine invariant, namely length ratio, to represent the connected components. Components in a document image are classified into a few clusters, each of which is assigned with a representative symbol. Retrieval are based on "words" comprising of symbols. The experiment results showed that the proposed method achieved an average retrieval precision of 93.43% and recall of 94.22%.