Auto-Labeled Image Sampling for Neural Network Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional methods for training deep learning networks are labor-intensive and costly due to the need for extensive manual labeling, with low inspector throughput and difficulty in acquiring sufficient skilled inspectors to match the auto-labeling device's pace, leading to inefficiencies in generating true labels.

Innovation Solution

A method and device that selectively acquire sample images for label-inspecting by transforming auto-labeled images using convolutional and pooling layers, identifying difficult images, and classifying them for efficient inspection by unskilled inspectors, thereby optimizing labeling throughput and reducing the number of inspectors required.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If auto-labeling is used to increase labeling throughput, then productivity is improved, but the number of inspectors required increases leading to increased cost

Engineering Contradiction:
Improvelabeling throughputVSAvoidnumber of inspectors
Core Design Contradiction:
ProductivityVSQuantity of substance

Solution Approach 1:

The patent segments the inspection workload by introducing an automated difficulty classification system that divides images into different categories (easy, medium, hard). This allows the system to automatically handle easy images while routing only difficult cases to human inspectors, thereby segmenting the work based on complexity rather than requiring all inspectors to review all images.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an automated difficulty classification model as an intermediary between the auto-labeling system and human inspectors. This intermediary automatically assesses image difficulty and routes appropriate images to inspectors, acting as a mediator that optimizes the workflow and reduces the burden on human inspectors.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Productivity

If more inspectors are hired to match auto-labeling throughput, then productivity is improved, but cost increases

Engineering Contradiction:
Improveinspection throughputVSAvoidnumber of inspectors
Core Design Contradiction:
ProductivityVSQuantity of substance

Solution Approach 1:

The patent changes the parameter of image difficulty assessment by introducing an automated classification system that evaluates images based on multiple features (object count, occlusion level, image quality metrics). This parameter change allows the system to dynamically determine which images require human inspection, optimizing the distribution of work without increasing inspector numbers.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If manual labeling is performed to ensure data quality, then measurement precision is improved, but time consumption increases

Engineering Contradiction:
Improvelabeling accuracyVSAvoidlabeling time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by having inspectors review and annotate difficult images first, before finalizing the entire dataset. The difficulty classification system identifies images that require human attention in advance, allowing inspectors to focus their efforts on cases most likely to contain errors while auto-labeling handles the majority of images.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements feedback mechanisms where inspector corrections are used to retrain and improve the auto-labeling and difficulty classification models. This feedback loop continuously improves labeling accuracy over time while maintaining efficient throughput, as the system learns from inspector interventions.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3690714B1Method for acquiring sample images for inspecting label among auto-labeled images to be used for learning of neural network and sample image acquiring device using the same
Publication Date: 2023.12.06 STRADVISION
  • EP3690714B1 patent drawingFigure 1
  • EP3690714B1 patent drawingFigure 2
  • EP3690714B1 patent drawingFigure 3

AI summary

A method for acquiring a sample image for label-inspecting among auto-labeled images for learning a deep learning network, optimizing sampling processes for manual labeling, and reducing annotation costs is provided. The method includes steps of: a sample image acquiring device, generating a first and a second images, instructing convolutional layers to generate a first and a second feature maps, instructing pooling layers to generate a first and a second pooled feature maps, and generating concatenated feature maps; instructing a deep learning classifier to acquire the concatenated feature maps, to thereby generate class information; and calculating probabilities of abnormal class elements in an abnormal class group, determining whether the auto-labeled image is a difficult image, and selecting the auto-labeled image as the sample image for label-inspecting. Further, the method can be performed by using a robust algorithm with multiple transform pairs. By the method, hazardous situations are detected more accurately.