Image Classification Training with Offset Networks for Semantic Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The high cost and low efficiency of manual pixel-level annotation in training semantic image segmentation networks, particularly for images with resolutions like 1024×2048, result in excessive manual annotation costs and reduced training efficiency.

Innovation Solution

A method for training an image classification model using an offset network and image classification network, where model parameters are adjusted through a classification loss function to reduce the need for manual pixel-level annotation, enabling the training of semantic image segmentation networks with image-level annotations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual pixel-level annotation is used for training semantic image segmentation networks, then the model can achieve accurate segmentation performance, but the annotation cost becomes excessively high and training efficiency decreases

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

Solution Approach 1:

The method performs preliminary full-image classification to generate initial segmentation regions before detailed pixel-level processing. By pre-identifying content regions through image-level classification, the system reduces the scope of subsequent pixel-level annotation work, thereby decreasing annotation time while preserving segmentation accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The annotation process is segmented into two stages: first, image-level classification to identify content regions; second, pixel-level processing only within those identified regions. This segmentation of the annotation task reduces overall annotation time while maintaining accuracy in the regions that matter most.

Inventive Principle:
Principle #1Segmentation

2Loss of information

If manual pixel-level annotation is performed on high-resolution images (e.g., 1024×2048), then complete segmentation information can be obtained, but the annotation cost becomes excessively high

Engineering Contradiction:
Improvesegmentation information completenessVSAvoidannotation cost
Core Design Contradiction:
Loss of informationVSLoss of time

Solution Approach 1:

The system applies different processing qualities to different regions: image-level classification is applied to the entire high-resolution image to identify content regions, while pixel-level processing is applied locally only to those identified regions. This local quality approach ensures complete segmentation information where needed while reducing overall annotation cost.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The method extracts and processes only the relevant content regions identified by the classification network, rather than processing the entire high-resolution image uniformly. By taking out and focusing on only the necessary regions, the system maintains information completeness while reducing annotation cost.

Inventive Principle:
Principle #2Taking out (Extraction)

3Difficulty of detecting and measuring

If full-image classification is used to locate image content regions, then the regions can be identified for segmentation, but the process still requires pixel-level annotation which reduces training efficiency

Engineering Contradiction:
Improvecontent region location accuracyVSAvoidtraining efficiency
Core Design Contradiction:
Difficulty of detecting and measuringVSProductivity

Solution Approach 1:

The method merges the content region location function with the segmentation training process. The classification network's content region detection capability is combined with the segmentation network training, allowing the system to both locate regions accurately and train the segmentation model simultaneously, thereby improving training efficiency without sacrificing location accuracy.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The classification network serves multiple functions: it locates image content regions, provides supervision signals for segmentation training, and identifies areas requiring pixel-level processing. This multi-functionality improves training efficiency by eliminating the need for separate processing steps while maintaining accurate content region detection.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3951654B1Image classification model training method, and image processing method and device
Publication Date: 2025.10.15 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3951654B1 patent drawingFigure 1
  • EP3951654B1 patent drawingFigure 2
  • EP3951654B1 patent drawingFigure 3

AI summary

Disclosed is an image classification model training method, comprising: obtaining an image to be trained; when a first model parameter of an offset network to be trained is fixed, obtaining first prediction category marking information of said image by means of an image classification network to be trained; determining a second model parameter using a classification loss function according to image content category information and the first prediction category marking information; when a second model parameter of said image classification network is fixed, obtaining second prediction category marking information of said image by means of said offset network; determining a third model parameter using the classification loss function according to the image content category information and the second prediction category marking information; and obtaining an image semantic segmentation network model according to the second model parameter and the third model parameter. The present application implements pixel-level marking and model training by means of AI, and manual marking is not needed, so that the manual marking cost is lowered, and the model training efficiency is improved.