Cameras are increasingly expected to operate in difficult visual conditions, from night-time roads and underground mines to unattended industrial sites, rainy streets, and remote outdoor monitoring systems. Yet in these scenarios, imaging systems often face the same technical challenge: insufficient brightness, heavy noise, motion blur, and unstable video quality, all of which can affect downstream AI recognition and safety-related decision-making.
Developed by the “Ruying Suixing” student innovation team from the Glasgow College, University of Electronic Science and Technology of China, NyxEye is designed to address this challenge.
NyxEye is a software-hardware integrated AI-ISP solution for ultra-low-light dynamic imaging. Instead of simply brightening dark images, it integrates lightweight AI algorithms with the traditional Image Signal Processing pipeline, aiming to upgrade night-time video from “visible” to clear, recognizable, and evidence-ready.
Conventional night-vision technologies often face three major trade-offs. First, strong denoising may remove important textures, edges, and text details, while weak denoising leaves visible grain and noise. Second, in dynamic scenes, traditional multi-frame denoising can introduce ghosting, motion blur, flicker, and temporal instability, especially around pedestrians and vehicles. Third, many end-to-end AI enhancement models perform well in laboratory settings but are difficult to deploy on low-power edge cameras due to high computational and memory requirements.
NyxEye takes a different engineering-oriented approach. Rather than replacing the existing ISP pipeline, it uses a lightweight AI model as an intelligent control module to dynamically schedule mature image-processing operators such as 2DNR and 3DNR.
This approach combines the reliability of traditional ISP systems with the scene-understanding capability of AI, offering a practical route toward deployable low-light video enhancement.
According to the project materials, the team has built a RAW-domain low-light video data foundation and completed prototype validation of its 2DNR/3DNR fusion strategy. NyxEye has achieved 36.65 dB PSNR and 0.9412 SSIM in RAW-domain evaluation. The model contains approximately 0.86M parameters, with the goal of supporting 30+ FPS real-time processing for 1080P video on mainstream edge SoC platforms.
The potential applications of NyxEye extend beyond visual enhancement. In smart mining and industrial inspection, it may improve the recognition of equipment abnormalities, unsafe human behavior, and foreign objects on conveyor belts. In intelligent driving, it can support better perception of pedestrians, vehicles, lanes, and obstacles at night. In public security and campus or hospital monitoring, it can improve the usability of night-time surveillance. In ecological protection, forest safety, maritime and river monitoring, and remote outdoor stations, it can enhance the visibility of weak-light and long-distance moving targets.
Many existing night-vision solutions depend on infrared illumination, large-aperture lenses, high-end sensors, or multi-camera systems, which can increase cost, power consumption, heat, and deployment complexity. NyxEye aims to provide a more lightweight and scalable alternative by enabling ordinary lenses, single sensors, and existing edge devices to achieve stronger night-time imaging performance through AI-ISP integration. From algorithm IP and SDK integration to AI-enhanced night-vision camera products, NyxEye is exploring a new path for ultra-low-light dynamic imaging.
Its goal is simple: to make visual perception more reliable when light is at its weakest.