Deep Neural Network QP Prediction for Visual Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional video encoding methods assign Quantization Parameters (QP) globally, leading to inconsistent visual quality across image blocks due to varying sensitivity to compression distortion, as different blocks have different abilities to conceal distortion caused by compression.
Innovation Solution
A deep neural network-based QP mapping method that uses texture-descriptive features like Haralick texture descriptors, total-variation, and variance, pre-trained with an auto-encoder framework, to predict target QP values, ensuring consistent visual quality by dynamically allocating bits to sensitive image blocks, employing a spatial pyramid framework for feature extraction at various granularities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If global QP assignment is used, then encoding simplicity is maintained, but visual quality consistency across image blocks deteriorates
Solution Approach 1:
The patent divides the video frame into multiple image blocks and processes each block independently with its own QP value. The system segments the global encoding task into local block-level decisions, allowing different QP values to be assigned to different regions based on their specific characteristics, thereby resolving the contradiction between simple global encoding and consistent local quality.
Solution Approach 2:
The patent implements local quality control by computing distortion sensitivity metrics for each image block and assigning QP values individually based on these local characteristics. Blocks with higher distortion sensitivity receive lower QP values (better quality), while less sensitive blocks receive higher QP values (lower quality), ensuring visual quality consistency across the entire frame despite varying local requirements.
2Device complexity
If uniform bit allocation is used, then encoding complexity is reduced, but distortion concealment ability across image blocks deteriorates
Solution Approach 1:
The patent dynamically changes the QP parameter for each image block based on computed distortion sensitivity metrics. Instead of using a fixed uniform QP value, the system adjusts the QP parameter locally to match the distortion concealment needs of each block, thereby improving reliability without significantly increasing overall encoding complexity.
Solution Approach 2:
The patent performs preliminary analysis of each image block to compute distortion sensitivity metrics before final QP assignment. This preliminary action allows the system to pre-determine which blocks need better protection, enabling more reliable distortion concealment while keeping the actual encoding process relatively simple.
3Manufacturing precision
If per-block QP optimization is implemented, then visual quality consistency is improved, but computational complexity increases
Solution Approach 1:
The patent implements a self-service mechanism where each image block automatically computes its own distortion sensitivity metric and determines its optimal QP value based on this self-assessment. This eliminates the need for complex centralized optimization algorithms, reducing computational complexity while maintaining per-block quality optimization.
Solution Approach 2:
The patent replaces complex mechanical optimization algorithms with a simpler metric-based approach. Instead of using computationally intensive optimization routines to determine QP values, the system substitutes a direct computation of distortion sensitivity metrics that naturally guide QP assignment, significantly reducing computational complexity while maintaining visual quality consistency.
Data Source
AI summary
A QP mapping method is capable of preserving consistent visual quality across the encoded frame. It automatically assigns more bits to image blocks which are more sensitive to compression distortion. The texture-descriptive features employed for QP prediction are fast to compute and, together with a deep neural network, are able to effectively approximate the underlying QP mapping strategy deduced from a visual quality measure.