Towards Physical Distortion Identification and Removal in Document Images

Physical distortions, next to digital artefacts, are commonly seen in document images. Their presence sabotages the optical character recognition (OCR) process which not only leads to a reduced amount of automatically retrievable content, but also deteriorates the performance of other document analysis algorithms that rely on layout analysis or content recognition. This paper proposes a method to identify and remove certain types of physical distortions from document images. By exploiting the intensity and spatial relation of distorted pixels, we construct a conditional random field (CRF) based method for distortion identification. Furthermore, a peak searching method is proposed so that the model parameters of the energy functions in the conditional probability are automatically learnt from the image. Discrimination of the pixels from original document content and those from physical noises is obtained by maximizing the conditional probability in the CRF model. Examples from real-life image samples demonstrate the effectiveness of the proposed method.