Adaptive document image skew estimation

The skew of the scanned document image is inevitable, and its correction improves the performance of document recognition systems. Skew specifies the text lines deviation from the horizontal or vertical axes. To date, skew estimation algorithms have employed specific features in a repetitive process. We can improve these algorithms by simply using an adaptive algorithm. This approach is suitable when we have large number of similar documents. This paper proposes adaptive document image skew estimation algorithm using the features of existing methods and supervised learning. This approach significantly improves the skew estimation time and accuracy. The time improvement comes from the training that need be performed only once on the training images rather than the repetitive process for each image of previous algorithms. The accuracy improvement comes from the appropriate selection of features, learning algorithm and image adaptively. This method works well in all skew ranges up to 0.1°