End-to-End Image Classification Training With Reference Outputs

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current image classification methods suffer from high calculation complexity due to independent processes for feature extraction and classification, leading to increased resource consumption and reduced efficiency.

Innovation Solution

An image classification model is trained end-to-end by constructing a reference classification result based on the model's output, adjusting parameters to minimize the total error value between the classification results and the reference, thereby reducing complexity and improving accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If image features are extracted by a neural network and then classified by an independent classification module, then the classification can be performed, but the calculation complexity is high and resource consumption increases

Engineering Contradiction:
Improveclassification accuracyVSAvoidcalculation complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges the feature extraction process and classification process into a single unified model. The model directly outputs classification results from input images without requiring separate feature extraction and classification modules, thereby reducing calculation complexity while maintaining classification accuracy.

Inventive Principle:
Principle #5Merging (Combining)

2Reliability

If image features are extracted by a neural network and then classified by an independent classification module, then the classification can be performed, but resource consumption increases

Engineering Contradiction:
Improveclassification accuracyVSAvoidresource consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent combines feature extraction and classification into one integrated model, eliminating the need for separate processing stages. This integration reduces computational overhead and resource consumption while preserving the ability to achieve accurate classification results.

Inventive Principle:
Principle #5Merging (Combining)

3Reliability

If image features are extracted by a neural network and then classified by an independent classification module, then the classification can be performed, but the classification efficiency is reduced

Engineering Contradiction:
Improveclassification accuracyVSAvoidclassification efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent integrates feature extraction and classification operations into a single model that processes images end-to-end. This merging eliminates the sequential processing bottleneck of separate modules, improving classification efficiency while maintaining accuracy.

Inventive Principle:
Principle #5Merging (Combining)

4Device complexity

If a unified model directly outputs classification results, then the complexity of the image classification process is lowered, but the model structure needs to be simplified

Engineering Contradiction:
Improveprocess complexityVSAvoidmodel structure design
Core Design Contradiction:
Device complexityVSEase of manufacture

Solution Approach 1:

The patent extracts and removes unnecessary complex components from the model structure. By eliminating the need for separate feature extraction modules and complex multi-stage processing, the model achieves a simplified architecture that is easier to design and implement while maintaining low process complexity.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4113376B1Image classification model training method and apparatus, computer device, and storage medium
Publication Date: 2025.11.19 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP4113376B1 patent drawingFigure 1
  • EP4113376B1 patent drawingFigure 2
  • EP4113376B1 patent drawingFigure 3

AI summary

This application discloses an image classification model training method and apparatus, a computer device, and a storage medium, which belong to the field of image processing technologies. In this application, classification results of each image outputted by an image classification model are obtained. When the classification results outputted by the image classification model do not meet a reference condition, a reference classification result is constructed based on the classification results outputted by the image classification model. Because the reference classification result can indicate a probability that images belong to each class, a parameter of the image classification model is updated to obtain a trained image classification model based on a total error value between the classification results of the each image and the reference classification result.