YolovN-CBi: A Lightweight and Efficient Architecture for Real-Time Detection of Small UAVs
This paper proposes YolovN-CBi, a lightweight and efficient architecture enhanced with CBAM and BiFPN modules, which, when combined with knowledge distillation, achieves superior speed and accuracy for real-time small UAV detection compared to baseline and newer YOLO versions.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the sky above our cities and countryside, a new kind of traffic is taking shape. Unmanned aerial vehicles, commonly known as drones, have become ubiquitous tools for everything from capturing aerial photography to monitoring crops and inspecting infrastructure. Their versatility and low cost have made them indispensable in modern life. Yet, this same accessibility has introduced a significant security challenge. Because they are small, quiet, and capable of rapid, unpredictable movement, drones can easily slip past traditional security measures. They can be used for unauthorized surveillance or even carry dangerous payloads, posing risks to sensitive areas and public safety. To counter this, security teams need systems that can spot these tiny flying objects instantly, even when they are far away and appear as mere specks against a complex background.
For years, researchers have relied on radar and sound to find these intruders, but these methods often struggle in busy, dynamic environments. In recent years, the focus has shifted to computer vision, where cameras and artificial intelligence work together to "see" and identify drones in real time. The most popular tools for this job are a family of AI models called YOLO, which stands for "You Only Look Once." These models are designed to scan an image and find objects quickly, making them ideal for live video feeds. However, a persistent problem has remained: while these models are excellent at finding large objects like cars or people, they often miss the smallest targets. When a drone is hundreds of meters away, it might occupy only a few pixels on a camera screen, and standard AI models tend to overlook such tiny details in their rush to process the rest of the image.
A team of researchers from the Homi Bhabha National Institute and the Bhabha Atomic Research Center in India has tackled this specific weakness. They set out to build a system that is not only fast enough to run on portable, battery-powered devices but also sharp enough to catch a drone that is barely visible. Their approach involved taking a proven, efficient AI architecture and carefully modifying its internal "eyes" to pay closer attention to the smallest details. Instead of simply adopting the newest, most complex versions of these AI models, which often require massive computers to run, they refined an earlier, lighter version. They added two specific upgrades to help the system focus. The first upgrade acts like a spotlight, teaching the model to ignore irrelevant background noise and concentrate on the most important parts of the image. The second upgrade improves how the model combines information from different distances, ensuring that tiny objects seen from far away are not lost as the image is processed.
The researchers tested their new design, which they named YolovN-CBi, against a wide range of existing models. They trained it on a massive collection of 28,000 images featuring various flying objects under different lighting conditions, including sunny days, cloudy skies, and evening light. To ensure the system was truly ready for the real world, they also created a local test set of 2,500 images taken from a fixed camera, capturing drones flying at a distance of about 350 meters. In these images, the drones were incredibly small, sometimes appearing as just 10 to 20 pixels. The results were striking. The researchers found that their modified model, built on an earlier version of the AI architecture, outperformed the latest and most advanced versions currently available. While the newer models were faster at identifying large objects, they frequently failed to detect the tiny drones in the distance. The new design, however, successfully identified these small targets with high accuracy, proving that a carefully tuned, lighter model can be more effective than a heavier, more complex one for this specific task.
To make this powerful system usable on small, portable devices like those mounted on security towers or carried by patrol officers, the team took a further step. They used a technique called knowledge distillation, which is similar to a master teacher guiding a student. They took the highly accurate, larger model they had built and taught a much smaller, faster version to mimic its decisions. This process allowed the smaller model to learn the subtle patterns of a distant drone without needing the heavy computational power of the original. The final result was a lightweight system that was nearly 83% faster than the original large model while maintaining, and in some cases even improving, its ability to spot the tiny intruders. This distilled model could process video frames in just a few milliseconds, making it fast enough for real-time surveillance.
The study also challenged a common assumption in the field of artificial intelligence: that the newest version of a tool is always the best. The researchers explicitly showed that for the difficult task of spotting small, distant drones, the latest models did not necessarily offer better performance. In fact, some of the newest versions were slower and less accurate at this specific job than their predecessors. By focusing on the specific needs of the task—sensitivity to tiny objects and the need for speed—the team demonstrated that a well-optimized, slightly older architecture could outperform the cutting-edge alternatives. Their work provides a clear path forward for security systems, showing that with the right adjustments, it is possible to build a lightweight, highly accurate detector that can run on the edge of the network, ready to protect airspace in real time.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.