Filterfusion enables fast and accurate video noise removal on CPUs
FilterFusion: Prior-Guided Adaptive Filter Fusion for Efficient Video Denoising
Computer Vision and Pattern Recognition
Summary
Videos can get noisy or blurry because of camera issues or compression, and cleaning them up is often slow or requires powerful computers. The authors created FilterFusion, a method that mixes several simple filters using a smart guide to clean noisy videos quickly. This approach uses very little computing power yet matches the quality of bigger, slower methods. It runs fast enough for real-time video calls on just one CPU core.
What this means in practice
- •For video codec engineers: Integrate fast denoising that adapts to multiple noise types into video communication systems under tight latency constraints.
- •For mobile app developers: Add efficient real-time video noise removal to video chat or streaming apps that run on low-power devices without dedicated GPUs.$Commercial implications: This paper enables selling apps with superior live video quality on affordable devices by dramatically reducing CPU requirements.
Authors
Wenlong Yang, Weihang Yuan, Canran Jin, Jiahao Li, Henrik Turbell, Lifeng Sun
Abstract
Real-world video denoising for communication systems must handle mixed degradations such as sensor noise, compression artifacts, and motion-induced inconsistency while remaining deployable under tight latency budgets. Purely deep models often achieve high accuracy but come with significant computational cost, whereas classical filters are robust and efficient but insufficiently adaptive. We present FilterFusion, a prior-guided multi-frame denoising framework that combines temporally-aware learned features with an interpretable bank of denoising branches. The model predicts a low-cost guidance signal and uses it to adaptively mix a compact set of denoising responses at the original resolution. The lightweight design uses only 109K parameters and explicitly injects temporal variance and spatial variance cues into the low-resolution controller. On CRVD (real sensor noise), it reaches 34.72 dB PSNR and 0.9330 SSIM with only 0.18 GFLOPs (network-only). Because the filter-bank operations are simple and regular (Gaussian blur, median, weighted sum), a straightforward compiled C implementation of the synthesis path achieves 17.9ms per frame (55.9 FPS) on a single CPU thread at 1080p---exceeding real-time and 122--252 times faster than lightweight CNN baselines that require over 2 seconds per frame. The results demonstrate a strong quality-efficiency-speed trade-off and a practical route to real-time denoising through low-resolution control and explicit filter-based synthesis.