Adaptive Quantization for CNN Video Filters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
NN-based filtering techniques in video encoding and decoding are complex and require significant processing power, with floating-point operations leading to reproducibility issues across hardware platforms and decoding drifts, and fixed quantization precision introduces unnecessary deviations between floating-point and integer representations.
Innovation Solution
Deriving a quantization value for a CNN filter layer based on NN weights and input properties to convert floating-point precision values to integer precision, reducing computational complexity and eliminating decoding drifts while maintaining coding quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point precision values are used for NN weights, then filtering quality is maintained, but computational complexity increases and decoding drift occurs across different hardware platforms
Solution Approach 1:
The patent changes the precision parameter from floating-point to integer by deriving a quantization value based on the maximum absolute value of NN weights. This parameter change reduces computational complexity while maintaining filtering quality through adaptive quantization that preserves important weight information.
Solution Approach 2:
The patent substitutes floating-point arithmetic operations with integer arithmetic operations. This replacement eliminates the computational complexity and hardware dependency issues associated with floating-point operations while maintaining the essential filtering functionality through quantized weight values.
2Device complexity
If fixed quantization precision is used to convert floating-point to integer values, then computational complexity is reduced, but deviation between floating-point and integer representations increases
Solution Approach 1:
The patent makes the quantization precision dynamic rather than fixed by deriving the quantization value from the actual NN weight distribution (maximum absolute value). This dynamic adaptation allows the quantization precision to match the specific characteristics of each layer, reducing representation deviation while maintaining computational efficiency.
Solution Approach 2:
The patent applies different quantization precision to different layers of the neural network based on their specific weight characteristics. Each layer derives its own quantization value, allowing locally optimized precision that reduces representation error for that specific layer while maintaining overall computational efficiency.
3Measurement precision
If increasing quantization precision is used to reduce deviation between floating-point and integer representations, then filtering quality improves, but the number of operations exceeding dynamic range increases requiring clipping
Solution Approach 1:
The patent performs preliminary scaling of NN weights by the quantization value before filtering operations. This preliminary action prevents overflow by ensuring that quantized weight values remain within the dynamic range, eliminating the need for clipping operations while maintaining filtering quality.
Data Source
AI summary
A device for decoding video data is configured to determine first neural network (NN) weights for a layer of a convolutional neural network (CNN) filter; derive a quantization value for the layer of the CNN filter based on values of the first NN weights; convert the first NN weights to second NN weights based on the quantization value; and filter a block of the video data using the second NN weights.