Learned Post-Processing Filter for Codec-Based Machine Vision

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing multimedia systems face challenges in optimizing the performance of machine vision tasks through traditional codecs, which often result in suboptimal compression and decoding efficiency.

Innovation Solution

Implementing a learned neural network-based filter that utilizes auxiliary information from traditional codecs to enhance the performance of machine vision tasks, particularly in video coding systems, by employing a post-processing filter trained on such information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If traditional codecs are used for data compression and decoding, then device complexity is reduced, but machine vision task performance deteriorates

Engineering Contradiction:

Improvecodec structureVSAvoidmachine vision task performance

Core Design Contradiction:

Device complexityVSMeasurement precision

Solution Approach 1:

A post-processing filter is introduced as an intermediary component between the traditional codec and the machine vision task. This filter takes the decoded output from the traditional codec and processes it further to enhance features relevant to machine vision tasks, thereby improving task performance without requiring changes to the traditional codec structure itself

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The system is segmented into distinct functional components: the traditional codec handles compression/decoding, while a separate post-processing filter handles machine vision optimization. This segmentation allows each component to specialize in its function without increasing overall system complexity

Inventive Principle:

Principle #1Segmentation

2Measurement precision

If neural networks are integrated into the codec, then machine vision task performance is improved, but device complexity increases

Engineering Contradiction:

Improvemachine vision task performanceVSAvoidcodec structure

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

Instead of fully replacing the traditional codec with a neural network-based codec, only a partial neural network component (the post-processing filter) is added. This partial application of neural networks provides the necessary performance improvement for machine vision tasks while avoiding the full complexity of end-to-end learned codecs

Inventive Principle:

Principle #16Partial or excessive action

3Loss of energy

If data compression is increased, then bitrate is reduced, but visual quality deteriorates

Engineering Contradiction:

ImprovebitrateVSAvoidvisual quality

Core Design Contradiction:

Loss of energyVSManufacturing precision

Solution Approach 1:

The post-processing filter converts the harmful effects of aggressive compression (artifacts, lost details) into beneficial outcomes by learning to reconstruct and enhance important features from the compressed data, effectively turning compression artifacts into opportunities for selective feature enhancement

Inventive Principle:

Principle #22Blessing in disguise (Convert harm into benefit)

Data Source

PatentUS20260113491A1Performance Improvements Of Machine Vision Tasks Via Learned Neural Network Based Filter

Publication Date: 2026.04.23 NOKIA TECHNOLOGIES OY

AI summary

A method includes receiving auxiliary information and/or at least one auxiliary feature, the at least one auxiliary feature being based on the auxiliary information; wherein the auxiliary information comprises information available within a decoder, or the auxiliary information comprises a bitstream output from data encoded using an encoder; receiving decoded data generated using the decoder; and generating filtered data with at least one filter using the auxiliary information and/or the at least one auxiliary feature via applying the filter to the decoded data; wherein the at least one filter comprises a learned filter; wherein the filtered data is configured to be used for at least one machine task performed using a model.