[Paper Review] A QP-adaptive Mechanism for CNN-based Filter in Video Coding

Chao Liu, Heming Sun+3 more|arXiv (Cornell University)|Oct 25, 2020

Video Coding and Compression Technologies21 references4 citations

TL;DR

This paper proposes a QP-adaptive mechanism that integrates quantization step (Qstep) directly into the convolutional kernel of CNN-based video in-loop filters, enabling a single model to generalize across multiple quantization parameters (QPs). By modeling quantization noise in the frequency domain and modifying the convolution operation to adaptively scale with Qstep, the method achieves performance comparable to multiple QP-specific models while reducing parameters by 75% and yielding an additional 0.2% BD-rate gain on chroma components.

ABSTRACT

Convolutional neural network (CNN)-based filters have achieved great success in video coding. However, in most previous works, individual models are needed for each quantization parameter (QP) band. This paper presents a generic method to help an arbitrary CNN-filter handle different quantization noise. We model the quantization noise problem and implement a feasible solution on CNN, which introduces the quantization step (Qstep) into the convolution. When the quantization noise increases, the ability of the CNN-filter to suppress noise improves accordingly. This method can be used directly to replace the (vanilla) convolution layer in any existing CNN-filters. By using only 25% of the parameters, the proposed method achieves better performance than using multiple models with VTM-6.3 anchor. Besides, an additional BD-rate reduction of 0.2% is achieved by our proposed method for chroma components.

Motivation & Objective

• To address the limitation of existing CNN-based in-loop filters that require separate models for each QP band, leading to high storage and computational costs.
• To improve the generalization capability of CNN filters across varying quantization noise levels without retraining for each QP.
• To leverage Qstep as a learnable input factor within the convolution operation, enhancing adaptability to different distortion levels.
• To maintain high performance with minimal increase in decoding complexity, enabling practical deployment.
• To outperform existing QP-adaptive methods that rely on QP maps only at the input layer, by integrating Qstep into every convolution layer.

Proposed method

• The method models quantization noise in the frequency domain, where increased QP corresponds to higher noise energy, and formulates a Qstep-injected convolution kernel to adaptively suppress noise.
• It modifies the standard convolution operation by incorporating Qstep into the kernel weights via a learned scaling factor, enabling dynamic noise suppression based on QP.
• The core equation introduces Qstep as a multiplicative factor in the convolution: $ \hat{\bm{x}} = \bm{w}(Q_{\text{step}}) \ast \bm{y} $, where the kernel $ \bm{w} $ is conditioned on Qstep.
• The mechanism is applied to all convolution layers (except the first), ensuring that Qstep influence propagates through the network, enhancing robustness.
• The method is plug-and-play, replacing standard convolution layers in any existing CNN filter without architectural changes.
• The approach is validated using VTM-6.3 anchor, training on luma only and evaluating on both luma and chroma components.

Experimental results

Research questions

• RQ1 Can a single CNN-based in-loop filter generalize effectively across multiple QP values without retraining?
• RQ2 How can Qstep information be effectively integrated into the convolution operation to improve noise suppression across varying QP levels?
• RQ3 Does injecting Qstep into the convolution kernel improve performance compared to QP map input at the first layer?
• RQ4 What is the trade-off between model complexity, parameter count, and rate-distortion performance when using a unified QP-adaptive model?
• RQ5 Can the proposed method achieve performance comparable to multiple QP-specific models while reducing parameter count and complexity?

Key findings

• The proposed method reduces the number of parameters by 75% compared to using separate models for each QP band, while achieving comparable BD-rate performance.
• The method achieves an additional 0.2% BD-rate reduction on chroma components, demonstrating superior generalization when trained only on luma.
• Performance remains consistently high across all QP values, with minimal degradation at low QP, unlike separate models that underperform outside their training QP range.
• The method outperforms the QP map input method (Song et al. [20]) in robustness, especially at low QP (e.g., QP=22), where the latter shows negative gains.
• The decoding complexity increases by only about 2% compared to the global method, indicating minimal overhead for practical deployment.
• The method maintains strong performance across four different CNN filter architectures of varying complexity, proving its versatility and generalization capability.