CNN Lens Defect Classification via Edge Surface Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep learning networks face challenges in processing images with sufficient accuracy and sensitivity, particularly in high-speed large-scale settings, and are not adapted to distinguish regions effectively in manufacturing operations for lens inspection systems.

Innovation Solution

The use of convolutional neural networks (CNNs), such as VGG16 Net and VGG19 Net, to separate lens edge and surface images into distinct datasets for analysis, with the option of using a single AI model to predict defects in both edge and surface images, and the application of class activation maps (CAMs) for model evaluation and correction.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If deep learning networks are used for lens image processing, then image interpretation capability is improved, but processing accuracy and sensitivity are insufficient for high-speed large-scale manufacturing settings

Engineering Contradiction:
Improveimage interpretation capabilityVSAvoiddefect detection accuracy
Core Design Contradiction:
Extent of automationVSMeasurement precision

Solution Approach 1:

The patent divides the lens inspection task into two separate AI models: one specialized for edge images and another for surface images. Each model is trained independently on its specific dataset, allowing them to develop specialized features for their respective domains. This segmentation enables each model to achieve higher accuracy in its specific task while maintaining automated processing capability.

Inventive Principle:
Principle #1Segmentation

2Extent of automation

If deep learning networks are used for lens image processing, then image interpretation capability is improved, but processing speed is insufficient for high-speed manufacturing settings

Engineering Contradiction:
Improveimage interpretation capabilityVSAvoidprocessing speed
Core Design Contradiction:
Extent of automationVSProductivity

Solution Approach 1:

By splitting the inspection system into two specialized models (edge model and surface model), each model becomes more efficient at its specific task. This specialization reduces computational complexity per model compared to a single general-purpose model, enabling faster processing speeds suitable for high-speed manufacturing environments while maintaining automated interpretation capability.

Inventive Principle:
Principle #1Segmentation

3Device complexity

If a single AI model is used for both edge and surface analysis, then device complexity is reduced, but the ability to distinguish between different regions is insufficient

Engineering Contradiction:
Improvemodel architecture complexityVSAvoidregion distinction capability
Core Design Contradiction:
Device complexityVSDifficulty of detecting and measuring

Solution Approach 1:

The patent implements separate AI models for edge and surface regions, with each model trained on dedicated datasets containing only that region type. This segmentation allows each model to develop specialized feature detection capabilities for its specific region, significantly improving the system's ability to distinguish between different lens regions and their respective defects, despite the increased complexity of having multiple models.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4352699B1Acquiring and inspecting images of ophthalmic lenses
Publication Date: 2024.10.16 COOPERVISION INT LTD
  • EP4352699B1 patent drawingFigure 1
  • EP4352699B1 patent drawingFigure 2
  • EP4352699B1 patent drawingFigure 3

AI summary

Systems and methods for acquiring and inspecting lens images of ophthalmic lenses using one or more cameras to acquire the images of the lenses in a dry state or a wet state. The images are preprocessed and then inputted into an artificial intelligence network, such as a convolutional neural network (CNN), to analyze and characterize for type of lens defects. The artificial intelligence network identifies defect regions on the images and output defect categories or classifications for each of the images based in part on the defect regions.