Fast and robust skew estimation in document images through bilinear filtering model

Skew estimation of scanned document is important for document analysis and recognition. Owing to the complexity inherent in the document, some methods only process the documents with a small skew angle or specified content, layout and have high computational cost. A fast and robust skew estimation method is proposed based on a bilinear filtering model, which is used to detect edges existing in the document. Some foreground areas in the document have been extracted without considering document layouts or contents. The proposed approach enhances the structure of the document and reduces the effects of the noises in the document. It combines filtering operators into a single approach, so noise filtering which is an unavoidable pre-processing in the previous skew detection methods has been overcome. A dominant angle has been estimated based on the detected edges. According to the estimated dominant angle, a skew angle can be determined efficiently without confining the search space or making assumptions including skew angle range and layout of the document in advance. The proposal greatly reduces the computational time and works in an unsupervised style. Comparative tests with the state-of-the-art skew estimation methods indicate the superior performance of the developed approach.