Image processing shifts from hardware to AI software, enabling real-time video improvement, updates, low-light detection, and performance on edge devices.
Chips&Media and Visionary.ai have introduced what they describe as the world’s first AI-based full Image Signal Processor (ISP), shifting image processing from fixed hardware pipelines to a software-defined system. The ISP uses neural networks to enhance video quality in real time.
At its core, the development replaces traditional ISP architectures that have remained largely unchanged for decades. Conventional ISPs are built as rigid, fixed-function hardware blocks, limiting image quality improvements and preventing updates after production. The new approach moves the entire imaging pipeline into software, where it can be tuned, optimized, and updated over the air.
The system builds on Visionary.ai’s earlier AI denoising technology, expanding it into a complete ISP pipeline. By combining AI models with Chips&Media’s hardware acceleration capabilities, the solution integrates multiple imaging functions into a single, adaptable framework.
In terms of performance, the AI ISP is designed to improve video quality across a wide range of lighting and motion conditions. It reduces blur and noise while increasing sharpness and color accuracy in real time. It also improves downstream computer vision tasks, with reported results showing over 75% improvement in object detection under low-light conditions and a 91% reduction in false positives.
The architecture is also focused on flexibility and deployment. Because the pipeline is software-defined, it allows instant tuning and continuous updates without requiring hardware changes. This enables devices to adapt to new conditions or use cases over time.
On the hardware side, Chips&Media addresses common AI trade-offs using its customized NPU, WAVE-N, developed from its experience in hardware IP design. The processor is designed to deliver high-precision, real-time imaging while maintaining low power consumption and competitive area efficiency.
To further optimize performance, the system uses line-by-line processing within CNN computation. This reduces DRAM bandwidth requirements and lowers both power consumption and latency, which are critical constraints in edge devices.
The ISP maintains these efficiency advantages while supporting 16-bit floating-point precision, enabling higher accuracy in image processing without increasing hardware overhead.
Click here for the original announcement.
