Canny2Palm: Realistic and Controllable Palmprint Generation for Large-scale Pre-training

TL;DR

This work tackles palmprint data scarcity for large-scale pre-training by introducing Canny2Palm, a palmprint synthesis framework that conditions a Pix2Pix generator on Canny edge textures and reassembles ROI blocks from multiple identities to create new, controllable identities. The method yields realistic palmprints and enables controllable intra-class diversity and inter-class distinction, obviating the need for a separate uniqueness classifier. Across open-set and cross-dataset evaluations, pre-training with Canny2Palm improves recognition performance, and the gains persist as synthetic IDs scale to 10k and beyond, including a million-scale evaluation. Ablation studies confirm the positive impact of background variation and border-cutout augmentation on diversity and generalization, underscoring the practical potential for large-scale palmprint pre-training in real-world deployments.

Abstract

Palmprint recognition is a secure and privacy-friendly method of biometric identification. One of the major challenges to improve palmprint recognition accuracy is the scarcity of palmprint data. Recently, a popular line of research revolves around the synthesis of virtual palmprints for large-scale pre-training purposes. In this paper, we propose a novel synthesis method named Canny2Palm that extracts palm textures with Canny edge detector and uses them to condition a Pix2Pix network for realistic palmprint generation. By re-assembling palmprint textures from different identities, we are able to create new identities by seeding the generator with new assemblies. Canny2Palm not only synthesizes realistic data following the distribution of real palmprints but also enables controllable diversity to generate large-scale new identities. On open-set palmprint recognition benchmarks, models pre-trained with Canny2Palm synthetic data outperform the state-of-the-art with up to 7.2% higher identification accuracy. Moreover, the performance of models pre-trained with Canny2Palm continues to improve given 10,000 synthetic IDs while those with existing methods already saturate, demonstrating the potential of our method for large-scale pre-training.

Figures (9)

• Figure 1: Example images of different synthetic datasets. (a) is the reproduced venous plexus of the palm based on the patterns of the Physarum; (b) is the vein texture generated by StyleGAN; (c) is the simulation of palmprint using B$\rm\acute{e}$zier curves, with a random background from ImageNet dataset; (d) and (e) are palmprint images with different gestures from the same identity based on our approach.
• Figure 2: The training procedure of the palmprint generator. The generator produces realistic palmprints based on the palm texture, while the discriminator tries to classify between the real and fake pairs.
• Figure 3: An illustration of the palm normalization and Canny edge extraction. (a) demonstrates 21 standard hand key-points; (b) is the normalized image by affine transformation with four reference joint points in red; (c) is the texture extracted by Canny.
• Figure 4: The synthesis procedure of a new identity. (a) illustrates $3\times3$ blocks within the ROI. (b) shows the cropped ROI on a new background. (c) is an example of the synthesis process. By replacing the blocks with corresponding areas from nine identities, a unique cropped ROI can be determined. The generator generates an unseen identity from this unique input. By replacing the background outside the same ROI, we generate new samples of the same identity for different gestures.
• Figure 5: Evaluation results on an internal million scale dataset.