Adaptive Block Switching Neural Prediction for Transform Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Adaptive block length encoding and decoding are incompatible with traditional decoding schemes due to varying block lengths, leading to inaccurate representation of media signals, and there is a need for improved methods to predict transform coefficients for adaptive block length media signals.

Innovation Solution

A neural network system is used to predict transform coefficients by converting quantized coefficients into enhanced coefficients, considering block length information and temporal/frequency dependencies, allowing for accurate representation of media signals.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional decoding schemes are used for adaptive block length signals, then the decoding process is simple, but the representation accuracy of media signals deteriorates

Engineering Contradiction:
Improverepresentation accuracyVSAvoiddecoding scheme complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

A neural network system is introduced as an intermediary between the quantized transform coefficients and the final decoded signal. The neural network predicts enhanced transform coefficients that compensate for quantization errors, thereby improving representation accuracy without requiring changes to the basic adaptive block length decoding framework

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The neural network dynamically adjusts the transform coefficients based on learned patterns from training data. By changing the parameters (coefficients) adaptively rather than using fixed traditional decoding methods, the system achieves higher accuracy in representing the original media signal

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If adaptive block length encoding is used, then the frequency and time resolution trade-off is optimized, but compatibility with traditional decoding schemes worsens

Engineering Contradiction:
Improvefrequency-time resolution optimizationVSAvoiddecoding compatibility
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The decoding process is segmented into distinct stages: first, traditional decoding reconstructs the quantized transform coefficients; second, the neural network processes these coefficients to predict enhanced values. This segmentation allows the system to maintain compatibility with traditional decoding schemes while adding adaptive enhancement capabilities

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The neural network is pre-trained on extensive media signal data before deployment. This preliminary training action enables the network to learn the statistical relationships and patterns in transform coefficients, so that during actual decoding, it can reliably predict enhanced coefficients that maintain compatibility with the adaptive block length encoding scheme

Inventive Principle:
Principle #10Preliminary action

3Productivity

If quantized transform coefficients are used directly, then the decoding process is fast, but the quality of decoded media signal deteriorates

Engineering Contradiction:
Improvedecoding speedVSAvoiddecoded signal quality
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

Instead of perfectly reconstructing all transform coefficients (which would be computationally expensive), the neural network applies partial enhancement by predicting only the most significant coefficient corrections. This excessive action beyond simple identity mapping provides quality improvement while maintaining relatively fast decoding performance

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The neural network creates a copy of the quantized transform coefficients and processes this copy to generate enhanced coefficients, while the original coefficients remain available for reference. This copying approach allows quality enhancement without fundamentally altering the fast decoding pipeline

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4229633B1Adaptive block switching with deep neural networks
Publication Date: 2025.08.06 DOLBY LABORATORIES LICENSING CORP
  • EP4229633B1 patent drawingFigure 1
  • EP4229633B1 patent drawingFigure 2
  • EP4229633B1 patent drawingFigure 3a~3b

AI summary

The present invention relates to a method for predicting transform coefficients representing frequency content of an adaptive block length media signal, by receiving a frame and receiving block length information indicating a number of quantized transform coefficients for each block in the frame, the number of quantized transform coefficients being one of a first or second number, wherein the first number is greater than the second number, determining a first block has the second number of quantized transform coefficients, converting the first block into a converted block having the first number of quantized transform coefficients, conditioning a main neural network trained to predict at least one output variable given at least one conditioning variable, the at least one conditioning variable being based on information regarding the converted block and block length information for the first block, providing at least one predicted transform coefficients from an output stage of the main neural network.