High Accuracy Fingerprint Recognition on FPGA Using a Dual Path Lightweight CNN

In the modern world, proving who you are often relies on something you carry, like a card, or something you remember, like a password. Yet these methods can be lost, stolen, or forgotten. This has led scientists and engineers to look for something you simply are, a biological trait that is unique to you and stays with you for life. Among these traits, the ridges and valleys on the tip of a human finger have long been a favorite for identification. They are easy to capture, remarkably stable over decades, and nearly impossible to duplicate perfectly. For years, computers have been taught to read these patterns, but doing so quickly and accurately without draining a battery or requiring a massive server has remained a difficult challenge. The goal is to build a system that can look at a fingerprint, understand its unique map, and make a decision in a flash, all while running on a small, efficient chip rather than a giant, power-hungry computer.

A team of researchers at the University of Florida has taken a significant step toward solving this problem by designing a new way for computers to recognize fingerprints that is both highly accurate and surprisingly light on resources. Instead of relying on the heavy, complex systems that usually power such tasks, they created a streamlined digital brain specifically built to fit onto a specialized computer chip known as a field-programmable gate array, or FPGA. These chips are like blank canvases for engineers; they can be programmed to perform specific tasks with incredible speed and very low energy use, making them ideal for devices that need to work reliably without plugging into a wall outlet. The researchers wanted to see if they could build a fingerprint reader that did not need to clean up or simplify the image before reading it, a common step that often throws away valuable detail.

To achieve this, the team designed a new type of neural network, a computer system modeled after the human brain that learns to recognize patterns by looking at thousands of examples. Most existing systems for this job are either too heavy to run on small chips or require the fingerprint image to be shrunk down and smoothed out, which can blur the very details needed for a correct match. The researchers proposed a different approach: a "dual-path" architecture. Imagine a team of two experts looking at the same fingerprint at the same time. One expert focuses on the big picture, looking at the overall shape and large curves of the print. The other expert zooms in to study the tiny, intricate details and fine lines. In their design, these two "paths" work in parallel, each extracting different kinds of information from the raw image. At the end, their findings are combined to make a final decision. This method allowed them to use the full, original size of the fingerprint images, preserving every ridge and valley without needing to resize or pre-process the data first.

The team trained this new system using a vast collection of fingerprint images from public databases, which included scans taken with different types of sensors and varying resolutions. To make sure the computer could handle the wide variety of real-world conditions, they taught it using a technique called data augmentation. This involved showing the system slightly altered versions of the same fingerprint, such as images that were rotated a few degrees or shifted slightly to the side, mimicking the way a finger might be placed differently on a scanner each time. By training on these varied examples, the system learned to recognize the underlying identity of the finger regardless of how it was positioned. When tested, this new model proved to be more accurate than several previous methods, correctly identifying fingerprints in about 70 percent of cases when using the full range of training data, and still achieving over 63 percent accuracy in stricter testing scenarios where the test images were not altered.

Beyond just being accurate, the true test of this research was whether it could actually run on the specialized hardware intended for real-world use. The researchers translated their digital design into a language that the FPGA chip could understand and built the entire system from the ground up. They managed to fit the complex network onto a Xilinx Virtex-7 FPGA, a powerful chip often used in high-performance applications. The result was a system that could process a single fingerprint in less than 80 milliseconds, fast enough for real-time use. The chip consumed about 29 watts of power, a figure that balances high performance with energy efficiency, and it utilized about 70 percent of the chip's available resources. This successful implementation demonstrates that it is possible to run sophisticated identity verification directly on efficient hardware, removing the need for massive external computers or heavy batteries.

The work also highlighted a crucial trade-off in how these systems are built. By reducing the number of internal settings, or parameters, that the network needs to remember, the researchers made the system much lighter and easier to store on a chip. Their model required only about 483 kilobytes of memory to hold all its learned information, a tiny fraction of what older, heavier models needed. This reduction in size did not come at the cost of performance; in fact, the new design outperformed previous attempts that used more complex structures. The researchers found that while one part of their dual-path system was better at spotting large features and the other at catching fine details, combining them was essential for the highest accuracy. This suggests that for fingerprint recognition, looking at both the forest and the trees simultaneously provides the most reliable answer.

Ultimately, this research offers a practical path forward for bringing secure, high-speed fingerprint recognition to devices that are small, portable, and energy-conscious. By proving that a lightweight, dual-path neural network can be built to run efficiently on an FPGA without sacrificing the details of the raw image, the team has shown that reliable biometric security does not require massive, power-hungry infrastructure. The system operates at a speed of 108.6 million cycles per second, making it fast enough to verify a person's identity almost instantly. This achievement suggests that in the near future, we may see fingerprint scanners embedded in a wider range of everyday devices, from mobile phones to border control gates, that are not only secure but also efficient and self-contained, capable of making complex decisions with a fraction of the power previously thought necessary.