Are Color Trained Models Robust in Handling Binary Images: A Fingerprint Recognition Study

Shubhi Gehlot, Poulomi Bhattacharya, Akshay Agarwal

Keywords: Fingerprint Recognition, Binary Images, Color Images, Deep Learning

Abstract: Fingerprint recognition has long been a cornerstone of biometric authentication, yet robust performance across varying imaging conditions remains a challenge, especially fingerphoto, which are generally acquired from the camera, instead of the Livescan images, which are not prone to the environmental factors. Due to the tremendous security demands in large-scale areas and areas where the deployment of computationally heavy devices might not be feasible, like refugee camps, the development of a scalable solution must be a priority. Through this research, we aim to achieve this by understanding the impact of binarization on images and models. Surprisingly, neither the role of Binarized Neural Networks (BNNs) nor binary fingerprint images (especially photos, not scans) has been explored in the literature. Henceforth, in this work, we conduct a comprehensive study of fingerprint recognition using both floating-point-based Deep Neural Networks (DNNs) and Binarised Neural Networks (BNNs) across multiple image representations, ranging from RGB to grayscale to binary. Our experiments reveal that while DNNs excel with richer representations such as RGB and grayscale, BNNs demonstrate superior compatibility with binary fingerprints, effectively leveraging their reduced complexity to achieve competitive or even better recognition accuracy. This finding highlights the importance of aligning model architectures with input spectra: full-precision networks benefit from information-rich domains, whereas binarized models coupled with binary images offer both efficiency and improved accuracy in inherently discrete representations. The results provide new insights into spectrum-aware fingerprint recognition, guiding the design of accurate and resource-efficient biometric systems.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 25417