Deep Learning In-Loop Filtering for QP-Adaptive Inter Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression techniques face challenges in managing varying levels of image distortion in P-frames and B-frames due to changing quantization parameter (QP) values, necessitating an adaptive in-loop filter to enhance coding efficiency.
Innovation Solution
A deep learning-based in-loop filter is employed to process P-frames and B-frames, utilizing an embedding vector and denoising model to mitigate image distortion by removing quantization noise, with components like a quantization parameter preprocessor and denoiser to enhance frame quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a conventional in-loop filter is used for inter-prediction, then the device complexity is low, but the coding efficiency is insufficient due to varying image distortion levels across different QP values
Solution Approach 1:
The patent implements a dynamic in-loop filter that adapts to varying QP values and image distortion levels. The filter's operation is controlled by a filter enable flag that is conditionally applied based on the current QP value and frame type (P-frame or B-frame), allowing the system to dynamically adjust filtering strength to match the distortion characteristics of different quantization parameters
Solution Approach 2:
The patent changes the filtering parameters (filter enable flag, filter strength) based on the quantization parameter value. Different QP ranges trigger different filtering behaviors, with higher QP values (greater distortion) enabling stronger filtering operations while lower QP values use minimal or no filtering, thereby optimizing coding efficiency across the full range of compression levels
2Manufacturing precision
If a deep learning-based in-loop filter is applied to mitigate image distortion, then the image quality is improved, but the use of energy and computational resources increases
Solution Approach 1:
The patent applies deep learning-based filtering selectively rather than uniformly across all frames. The filter is applied locally to specific frames (P-frames and B-frames) where distortion mitigation is most beneficial, determined by the filter enable flag condition. This localized application reduces overall computational energy while maintaining image quality where it matters most
Solution Approach 2:
The patent uses partial action by conditionally applying the deep learning filter only when needed (based on QP value and frame type) rather than applying it to every frame. This selective application provides sufficient distortion mitigation for high-QP frames while avoiding unnecessary computational overhead for low-QP frames, optimizing the energy-quality tradeoff
3Adaptability or versatility
If an adaptive in-loop filter is used to handle varying QP values, then the adaptability is improved, but the device complexity increases due to additional processing components
Solution Approach 1:
The patent implements a universal in-loop filter design that handles multiple frame types (P-frames and B-frames) and multiple QP value ranges using a single filter structure. The filter enable flag mechanism provides a unified control approach that adapts the same filter to different conditions, avoiding the need for separate specialized filters for each frame type or QP range, thereby managing complexity while maintaining versatility
Data Source
AI summary
A method and an apparatus for video coding using a deep learning-based in-loop filter for inter-prediction are disclosed. The video coding method and the apparatus utilize a deep learning-based in-loop filter for inter-prediction of a predictive frame (P-frame) and a bi-predictive frame (B-frame) in order to mitigate various levels of image distortion according to a QP (quantization parameter) value present in the P-frame and the B-frame.