Rotation, Scaling and Translation Analysis of Biometric Signature Templates

Biometric authentication systems that make use of signature verification methods often render optimum performance only under limited and restricted conditions. Such methods utilize several training samples so as to achieve high accuracy. Moreover, several constraints are imposed on the end-user so that the system may work optimally, and as expected. For example, the user is made to sign within a small box, in order to limit their signature to a predefined set of dimensions, thus eliminating scaling. Moreover, the angular rotation with respect to the referenced signature that will be inadvertently introduced as human error, hampers performance of biometric signature verification systems. To eliminate this, traditionally, a user is asked to sign exactly on top of a reference line. In this paper, we propose a robust system that optimizes the signature obtained from the user for a large range of variation in Rotation-Scaling-Translation (RST) and resolves these error parameters in the user signature according to the reference signature stored in the database.

💡 Research Summary

The paper addresses a fundamental limitation of offline signature‑based biometric authentication systems: their sensitivity to geometric variations introduced by users during signing. Traditional approaches try to mitigate these issues by constraining the user to a small signing box, forcing a fixed scale, and requiring the signature to be placed precisely on a reference line to limit rotation. While these constraints improve matching accuracy, they severely restrict usability and are still insufficient in real‑world scenarios where users naturally introduce rotation, scaling, and translation (collectively referred to as RST) errors.

The authors propose a complete preprocessing pipeline that automatically compensates for RST distortions before any feature extraction or matching takes place. The system uses a standard digital pen (Wacom Bamboo) to capture the signature as a raster image. After acquisition, the image is converted to grayscale and binarized, producing a clean black‑and‑white representation of the signature.

Rotation correction is performed by computing the normalized cross‑correlation (NCC) between the reference template and the user’s signature over a range of angles. An initial coarse search scans from –60° to +60° in 5° steps, locating the angle that yields the highest NCC value. A subsequent fine search refines this estimate within ±3° of the coarse result using 1° increments. The angle corresponding to the maximum NCC is taken as the optimal rotation; the user image is then rotated by the negative of this angle, aligning it with the reference. Normalization of the correlation makes the method robust to illumination and contrast differences, while the two‑stage search balances computational cost and precision.

Translation correction is achieved through a simple cropping operation. After binarization, the algorithm trims away all background pixels, leaving only the signature strokes. The lower‑left corner of the cropped region is defined as the new origin (0,0). Because both the reference and the user images undergo the same cropping, any translational offset—whether the user started signing near the left edge, the right edge, or anywhere in between—is eliminated without the need for explicit translation matrices.

Scaling correction follows the same cropping step. The height (and optionally width) of the cropped reference and user images are measured, and a scaling factor is computed as the ratio of reference size to user size. In the experiments, vertical scaling (height) exhibited more variation, so the Y‑scale factor was primarily used. The user image is resized according to this factor, bringing both images to the same pixel dimensions. This uniform scaling enables direct point‑wise comparison in the subsequent feature extraction stage.

After RST normalization, the system extracts geometric features from the aligned signature—typically a set of key points such as start, end, curvature extrema, and crossing points. Matching is performed by counting the number of corresponding points between the test signature and the stored template, producing a similarity score that is compared against a threshold to accept or reject the claim.

The experimental protocol involved 90 participants, each providing one reference signature and nine test signatures. The test set was deliberately distorted with rotations ranging from –60° to +60°, scaling factors between 0.5 and 2.0, and translations up to the full width or height of the signing canvas. Results show a substantial reduction in both false‑accept and false‑reject rates after applying the RST correction. For rotations within ±30°, the verification success rate exceeded 95 %; even with scaling up to 1.5×, accuracy remained above 90 %. Importantly, the method requires only a single reference sample per user, eliminating the need for large training databases.

The authors acknowledge several limitations. NCC‑based rotation estimation can degrade under heavy noise or severe non‑rigid deformation. The scaling step assumes isotropic scaling, making the approach vulnerable to anisotropic distortions where width and height change independently. Moreover, the robustness of the pipeline against varying scan resolutions, compression artifacts, or different acquisition devices (e.g., smartphone cameras) was not fully explored.

Future work is suggested in three directions: (1) implementing a multi‑scale, multi‑orientation search using parallel processing to improve robustness; (2) incorporating non‑linear deformation models such as Thin‑Plate Splines to handle anisotropic and elastic distortions; and (3) integrating deep‑learning‑based feature extractors (CNNs or graph neural networks) with the RST‑corrected images to create a hybrid system that leverages both geometric normalization and learned representations. Extending the approach to low‑cost hardware and evaluating cross‑device generalization are also highlighted as promising avenues.

In conclusion, the paper delivers a practical, low‑cost solution for RST‑invariant offline signature verification. By inserting a lightweight preprocessing layer that automatically aligns, crops, and rescales signatures, the system achieves high verification accuracy without imposing restrictive signing conditions on users. This contribution is significant for real‑world biometric deployments where convenience, scalability, and security must coexist.