Text Skew Angle Detection in Vision-Based Scanning of Nutrition Labels

An algorithm is presented for text skew angle detection in vision-based scanning of nutrition labels on grocery packages. The algorithm takes a nutrition label image and applies several iterations of the 2D Haar Wavelet Transform (2D HWT) to downsample the image and to compute the horizontal, vertical, and diagonal change matrices. The values of these matrices are binarized and combined into a result set of 2D change points. The convex hull algorithm is applied to this set to find a minimum area rectangle containing all text pixels. The text skew angle is computed as the rotation angle of the minimum area rectangle found by the convex hull algorithm. The algorithm’s performance is compared with the performance of the algorithms of Postl and Hull, two text skew angle algorithms frequently cited in the literature, on a sample of 607 nutrition label images whose text skew angles were manually computed by two human evaluators. The median text skew angle error of the proposed algorithm, Postl’s algorithm, and Hull’s algorithm are 4.62, 68.85, and 20.92, respectively. Keywords— computer vision; text skew angle detection; OCR; 2D Haar wavelet transform; wavelet analysis