An Iterative Decoding Approach to Document Image Analysis

We introduce an iterative approach to recognizing two-dimensional grammatical structure within digital images, which we term “turbo recognition.” Inspired by the success of turbo decoding for channel coding of one-dimensional sequences, we develop a recognition scheme for images based on two independent views of the same underlying message. These correspond to two independent image sources, one in the horizontal direction and the other in the vertical direction, which are driven by a single input message. The recognition process proceeds iteratively, first along one direction and then the other, applying the Forward/Backward algorithm to derive a new prior probability distribution on the input message for the orthogonal recognition step. This holds promise as a principled approach within the Document Image Decoding (DID) framework for the recognition of nontrivial 2D layout structure such as tables.