Neural Image Inspection with Pretrained Defect Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI image sensors for workpiece inspection require significant time and effort to establish learning-based inspection modes, necessitating retraining for different production lines, limiting usability.
Innovation Solution
An image inspection apparatus with a built-in illuminator, camera, and learned neural network storage that prepares and stores neural networks for failure/no-failure determination and classification, allowing for efficient inference without retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a learning-based inspection mode is established using a neural network, then inspection accuracy and adaptability are improved, but the time and effort required for setup and retraining increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-training neural networks for multiple different inspection scenarios, production lines, and product types before actual use. These pre-trained neural networks are stored in the storage unit, so when inspection is needed, the system can directly retrieve and use the appropriate pre-trained model without performing time-consuming training from scratch. This resolves the contradiction by preparing inspection models in advance, making the actual inspection process fast and adaptable while the training time is invested beforehand.
Solution Approach 2:
The patent implements universality by creating a single inspection apparatus that can perform multiple different types of inspections using different pre-trained neural networks. The system stores multiple neural networks that can handle various product types, production lines, and inspection criteria. By selecting the appropriate pre-trained network from storage, the apparatus becomes universally applicable across different inspection scenarios without requiring separate dedicated systems for each case, thus improving adaptability without proportionally increasing setup time.
2Measurement precision
If multiple dedicated neural networks are trained for different production lines and inspection types, then inspection precision is improved, but device complexity and training resources increase
Solution Approach 1:
The patent applies merging by combining multiple different neural networks into a single inspection apparatus. Instead of having separate dedicated systems for each production line or inspection type, the system integrates multiple pre-trained neural networks into one apparatus with a common hardware platform, processing unit, and storage unit. This reduces device complexity compared to having separate systems while maintaining high inspection precision for each specific task by selecting the appropriate pre-trained model.
Solution Approach 2:
The patent uses copying by creating multiple copies of neural network models (different pre-trained networks for different scenarios) and storing them in the storage unit. Rather than physically duplicating entire inspection systems, the approach copies the learned knowledge in the form of neural network parameters and weights. This allows the system to maintain multiple specialized inspection capabilities without proportionally increasing hardware complexity, as all networks share the same underlying infrastructure.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Facilitates quick and efficient failure/no-failure determination and classification of workpieces across different environments, reducing the need for extensive retraining and setup time.
Implementation Method 1
an illuminator that irradiates a workpiece as an inspection object with illumination light
Implementation Method 2
a camera that receives light that is reflected from the workpiece, which is irradiated by the illuminator, and produces a workpiece image
Data Source
AI summary
An image inspection apparatus includes a learned neural network storage storing a neural network that previously learns weighting factors between input, intermediate and output layers, and an inferer determining failure/no-failure of a workpiece and classify the workpiece to classes based on an image of the workpiece. The inferer performs first and second inferences. In the first inference, the inferer determines failure/no-failure of the workpiece based on failure/no-failure feature quantities that are obtained by providing the workpiece image to the neural network and a failure/no-failure determination boundary. In the second inference, the inferer define a classification boundary to be used to classify an inspection workpiece to the classes in a feature quantity space of the neural network based on classification feature quantities that represent the different-type classification workpiece images, and classifies a workpiece to the classes based on classification feature quantities of an image of the workpiece and the classification boundary.