Noise reduction is a key issue in any camera system to improve the visual appearance of the images. As the light level decreases, the sensor gain goes up and the noise level increases. AI 2D+3DNR is a deep-learning noise reduction IP core for very low light conditions, offered as a separate IP. It removes both spatial and temporal noise, including color noise and luminance noise, and keeps edges and fine texture. The deep-learning motion compensation aligns the previous frame with the current frame before temporal filtering, so moving objects keep their detail and ghost images are reduced. The strength follows the sensor gain, so the core works from normal light to very low light without manual tuning.
| Features |
|---|
| Deep-learning spatial noise reduction + temporal noise reduction |
| Deep-learning motion compensation for temporal filtering (fewer ghost images) |
| Noise reduction strength set automatically by the sensor gain |
| Bayer domain processing |
| Verified at 1920x1080 30 fps on AMD Kria KV260 (Zynq UltraScale+ MPSoC) with Sony IMX678 |
| Needs external DDR SDRAM for frame buffers |
The videos below compare the noisy input, our 2D+3DNR core and the AI 2D+3DNR core at several sensor gains. Pick a scene and a sensor gain, choose what to show on each side, and move the mouse over the video to move the divider.
Please view these videos on a PC with Chrome, Edge or Firefox. To show the noise exactly as it is, the videos are not compressed much (up to about 90 Mbps for the noisy input), so they may load slowly or not play on smartphones, tablets and in Safari.
Sensor gain = how much the sensor signal is amplified to make a dark scene look normal. +6 dB = ×2. 42 dB = ×126. Higher gain → darker scene → more noise.