Video Image Segmentation Using Reused Semantic Features

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current deep learning-based image segmentation methods require significant computational resources, limiting their application in real-time scenarios such as video calls or video conferences due to the high computational demands.

Innovation Solution

An image processing method that reuses the semantic feature of a previous image to predict the semantic feature of subsequent images in a sequence, reducing the need for re-extraction by using a noise reduction network, and employing a similarity threshold to ensure accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If deep learning-based image segmentation is performed on each captured image in real-time, then image segmentation accuracy is improved, but computational resource consumption increases significantly

Engineering Contradiction:
Improveimage segmentation accuracyVSAvoidcomputational resource consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent extracts semantic features from the first image in advance using a complete deep learning model, and then reuses these extracted features for subsequent images through a noise reduction network. This preliminary extraction and reuse mechanism avoids repeated full-model inference, significantly reducing computational resource consumption while maintaining segmentation accuracy for images with similar content.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If semantic feature extraction is performed on every image in the sequence, then segmentation accuracy is maintained, but processing time increases

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent creates a copy of the semantic features extracted from the first image and reuses this copied feature representation for subsequent images. By copying and adapting the semantic features through the noise reduction network instead of re-extracting them, the system significantly reduces processing time while maintaining segmentation accuracy for temporally adjacent images.

Inventive Principle:
Principle #26Copying

3Measurement precision

If the complete deep learning model is applied to all images, then segmentation quality is ensured, but device complexity requirements increase

Engineering Contradiction:
Improvesegmentation qualityVSAvoidcomputing resource requirements
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the image processing task into two parts: (1) complete semantic feature extraction using the deep learning model for the first image, and (2) feature refinement using a lightweight noise reduction network for subsequent images. This segmentation allows the complex model to be used only when necessary, reducing overall device complexity requirements while maintaining segmentation quality.

Inventive Principle:
Principle #1Segmentation

4Productivity

If feature reusing is implemented for subsequent images, then computational load is reduced, but similarity requirement between images must be met

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidsimilarity requirement
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent changes the parameter of image similarity by focusing on temporally adjacent images in a video sequence, which naturally have higher similarity due to continuous capture. This parameter change enables effective feature reuse and computational load reduction, as the noise reduction network can successfully adapt features from the first image to subsequent images that meet the similarity requirement.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4708215A1Image processing method, model training method and related apparatus
Publication Date: 2026.03.11 HUAWEI TECH CO LTD
  • EP4708215A1 patent drawingFigure 1
  • EP4708215A1 patent drawingFigure 2
  • EP4708215A1 patent drawingFigure 3

AI summary

An image processing method is provided, and is applied to the field of artificial intelligence technologies. In the method, continuously captured images in an image sequence are classified into a reference frame and a non-reference frame. For a non-reference frame in the image sequence, a semantic feature of a reference frame located before the non-reference frame is reused to predict a semantic feature of the non-reference frame, the semantic feature of the non-reference frame is no longer re-extracted, and then an image segmentation result of the non-reference frame is obtained through prediction based on the semantic feature of the non-reference frame. A similarity between the reference frame and the non-reference frame is effectively used in this solution, and a semantic feature extraction process for the non-reference frame is reduced by reusing the semantic feature of the reference frame. Therefore, an amount of computation of an image segmentation process for the non-reference frame can be effectively reduced, and an application scenario of the image processing method can be extended.