Compressed-Domain Inference With Dual Power Mode Processing

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Processing compressed image and video data in the pixel domain consumes excessive power and resources, and direct processing from Discrete Cosine Transform (DCT) coefficients lacks meaningful interpretation due to intra prediction being performed before transformation, leading to inaccurate reflection of image contents.

Innovation Solution

A system operates in two power modes: a first power mode for processing in the compressed domain by partly decoding and performing feature extraction on self-contained frames, and a second power mode for processing in the pixel domain by fully decompressing data, using neural networks with mode-specific models and parameters.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If full decompression is performed to process data in the pixel domain, then accurate inference processing is achieved, but power and resource consumption increases

Engineering Contradiction:

Improveinference accuracyVSAvoidpower consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system dynamically switches between two power modes (first power mode with partial decoding and second power mode with full decompression) based on operational requirements. This dynamic adaptation allows the system to optimize the balance between inference accuracy and power consumption by selecting the appropriate processing depth for each scenario.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

In the first power mode, the system performs partial decoding (only entropy decoding) rather than full decompression, extracting features directly from compressed data. This partial action approach achieves acceptable inference accuracy while significantly reducing power consumption compared to complete decompression.

Inventive Principle:

Principle #16Partial or excessive action

2Use of energy by moving object

If partial decoding is performed in the first power mode, then power consumption is reduced, but processing accuracy may be compromised

Engineering Contradiction:

Improvepower consumptionVSAvoidinference accuracy

Core Design Contradiction:

Use of energy by moving objectVSMeasurement precision

Solution Approach 1:

The system introduces an intermediary approach by performing entropy decoding to obtain transform coefficients and prediction modes without completing the full inverse transform and quantization processes. This intermediate state provides sufficient structural information for feature extraction while avoiding the computational cost of complete decompression.

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The system changes the processing parameters by adjusting the decoding depth according to power mode. In the first power mode, it uses reduced decoding parameters (entropy decoding only) to conserve energy, while in the second power mode, it uses full decoding parameters to maximize accuracy when power is abundant.

Inventive Principle:

Principle #35Parameter changes

3Use of energy by moving object

If DCT coefficients are processed directly, then power consumption is reduced, but meaningful interpretation is lost due to prior intra prediction transformation

Engineering Contradiction:

Improvepower consumptionVSAvoidimage content interpretation

Core Design Contradiction:

Use of energy by moving objectVSLoss of information

Solution Approach 1:

The system extracts only the essential structural information (transform coefficients and prediction modes) from the compressed data without performing the full decompression process. This extraction approach retrieves the key geometric and frequency characteristics needed for inference while avoiding the power-intensive steps that would restore complete image content.

Inventive Principle:

Principle #2Taking out (Extraction)

Data Source

PatentUS12505580B2Inference processing of data

Publication Date: 2025.12.23 TELEFONAKTIEBOLAGET LM ERICSSON (PUBL)

AI summary

A method for processing data in a system configured to operate in either of at least a first power mode and a second power mode, wherein the first power mode is associated with a first power level and the second power mode is associated with a second power level, the second power level being higher than the first power level, wherein the first and second power modes each are configured to prepare a respective model for inference processing is disclosed. The method comprises acquiring (101) compressed data, determining (102) whether the system operates in the first power mode or in the second power mode. The method further comprises, when the system operates in the first power mode, determining (103) whether the acquired compressed data comprises a self-contained frame, and if so partly decoding (104) the self-contained frame, performing (105) feature extraction of the decoded self-contained frame, preparing (107) the model for inference processing in the first power mode in the system, wherein the model comprises inference parameters for the first power mode, and performing (108) inference processing by a neural network based on the extracted features and the prepared model for inference processing. Corresponding computer program product, apparatus, and system are also disclosed.