Fixed-Length Rotation Invariance Representation for Fingerprint

Although convolutional neural network(CNN) based fixed-length fingerprint feature representations have achieved efficient and accurate matching performance, their effectiveness heavily relies on accurate fingerprint alignment. Existing fingerprint alignment methods transform fingerprints based on the predicted reference point coordinates and orientation. However, when dealing with arbitrarily rotated fingerprints, these methods struggle to predict the correct orientation, leading to rotation errors during alignment. Additionally, CNN exhibit relatively poor rotation invariance, making it difficult to extract consistent features from fingerprints with incorrect rotations. To address these issues, we propose a fixed-length, rotation invariant representation of fingerprint (RIRF), which is the first method to introduce polar coordinate mapping into CNN-based fingerprint feature extraction. The introduction of polar coordinate mapping bridges the rotational sensitivity of CNN and the rotation error of fingerprint alignment, effectively addressing the limitations of both and enhancing overall fingerprint recognition performance. Specifically, polar coordinate mapping transforms the fingerprint image into polar coordinate system, converting angular differences between various fingerprint acquisitions into translation variations in the image. Combining this with the translation invariance of CNN allows the extracted features to exhibit rotation invariance. Additionally, we introduce H-Circular Padding and Wide-wise Pooling to further enhance the model's invariance. Finally, experimental results show that RIRF consistently achieves strong recognition performance, even for arbitrarily rotated fingerprints, demonstrating superior rotation robustness compared to other methods.