Image Recognition Model Training via Confidence-Based Sample Selection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image recognition models require a large number of manually tagged samples for training, leading to low training efficiency and high labor costs due to the need for complete manual tagging.

Innovation Solution

A method that uses an initial small set of tagged positive and negative image samples to train an image recognition model, selecting samples with confidence levels within a preset interval for manual tagging, reducing the number of samples needed and enhancing the model's recognition capability.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If completely manual tagging is adopted for all image samples, then the image recognition model achieves adequate recognition capability, but the training efficiency is low and labor cost is high

Engineering Contradiction:
Improverecognition capabilityVSAvoidtraining efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The system enables the image recognition model to automatically evaluate its own confidence levels on untagged samples and self-select which samples need manual tagging, reducing dependency on complete manual tagging while maintaining recognition capability

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system uses confidence levels as feedback to determine which samples require manual tagging, creating a closed-loop process where the model's uncertainty guides the tagging effort allocation

Inventive Principle:
Principle #23Feedback

2Reliability

If completely manual tagging is adopted for all image samples, then the image recognition model achieves adequate recognition capability, but the labor cost is high

Engineering Contradiction:
Improverecognition capabilityVSAvoidlabor cost
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

Instead of tagging all samples, the system tags only the necessary portion of samples that fall within the confidence interval, achieving adequate recognition capability with reduced manual labeling effort

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The system changes the parameter from complete manual tagging to selective tagging based on confidence level thresholds, reducing the quantity of manual work required

Inventive Principle:
Principle #35Parameter changes

3Reliability

If a large number of tagged image samples are used for training, then the image recognition model achieves high accuracy, but the training time is increased

Engineering Contradiction:
Improverecognition accuracyVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The system uses only the necessary number of tagged samples (those within the confidence interval) for training, avoiding the time cost of training on all possible samples while maintaining high accuracy

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3855360B1Method and device for training image recognition model, and storage medium
Publication Date: 2024.09.04 BEIJING XIAOMI PINECONE ELECTRONICS CO LTD
  • EP3855360B1 patent drawingFigure 1~2
  • EP3855360B1 patent drawingFigure 3A~3B
  • EP3855360B1 patent drawingFigure 4

AI summary

A method and device for training an image recognition model, and a storage medium are provided. The method includes that: an image recognition model is trained by using a tagged image sample set; multiple to-be-tagged image samples in a to-be-tagged image sample set are recognized by using the presently trained image recognition model to obtain a confidence of a recognition result corresponding to each of the multiple to-be-tagged image samples; at least one to-be-tagged image sample of which the confidence is in a preset interval is selected as at least one to-be-tagged image sample to form a first to-be-tagged image sample set; tags of to-be-tagged image samples in the first to-be-tagged image sample set is acquired; the tagged image sample set is updated according to the acquired tags; and the image recognition model is trained by using the updated tagged image sample set.