Offline Signature Verification Based on Hu Moments

Abstract

This paper presents an offline signature verification system based on Hu moment invariants, implemented in MATLAB and evaluated on the publicly available CEDAR benchmark database. Offline signature verification operates solely on static images of completed signatures, without access to dynamic information such as pen pressure or writing speed, making it a challenging yet practically important problem for authenticating scanned documents and archival records. The proposed system follows a four-stage pipeline: image preprocessing, Hu moment-based feature extraction, exponentially weighted similarity scoring, and adaptive per-user decision making. Preprocessing includes grayscale conversion, intensity inversion, median and Wiener filtering, adaptive binarization, morphological cleaning, bounding-box cropping, and proportional resizing to a fixed 200x480 pixel canvas. Feature extraction employs the first five Hu moment invariants computed from normalized central moments, with non-uniform weighting favoring lower-order, more stable moments. Similarity is measured as an exponentially transformed weighted Euclidean distance, aggregated across reference templates with exponential weighting. The decision threshold is set adaptively per writer from leave-one-out intra-class similarity statistics, personalizing the security boundary to each individual's signing style. Experiments on all 55 CEDAR writers show that with five reference samples the system achieves 87.1% accuracy, 12.7% FAR, 13.1% FRR, and 12.8% EER, with an AUC of 0.940. These results confirm that Hu moment invariants are effective global shape descriptors for offline signature verification when paired with robust preprocessing and personalized adaptive thresholds.