A novel deep learning based method for Vietnamese ID card skew correction

The skewness introduced during the capturing or scanning of national identity cards poses a significant challenge of accurate information extraction. In this paper, we propose a method for skew estimations and corrections specifically for Vietnamese ID cards by leveraging deep learning techniques. Firstly, YOLOv8 Instance Segmentation is employed to generate the coordinates of the contour surrounding the identification card. Subsequently, a minimum area rectangle, covering the ID card segmentation mask, is created. The skew angle is then determined by the long side of the rectangle in relation to the image’s horizontal line. Next, the ID card is rotated based on the previously calculated skew angle. In the final step, the straightened image is passed through a classification model for upside-down correction. The method was evaluated on our comprehensive, self-developed datasets, showcasing its ability to accurately detect the rotation angle for skew ID cards across a broad range of angles.