Image Inspection Apparatus Using Feature Space Classification Boundary Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image inspection apparatuses face challenges in accurately distinguishing between non-defective and defective product images due to variations in imaging conditions and features, leading to unstable determination results, especially for complex objects like metal components.
Innovation Solution
An image inspection apparatus with a multi-layered learning processor that plots non-defective and defective product images in a feature space, generates a classifier, and estimates the probability of incorrect attributes for newly input images, allowing users to correct classifications and improve accuracy by adjusting the classification boundary.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image inspection apparatuses use threshold comparison with extracted features, then the determination process is simple and fast, but the accuracy is low and determination results are unstable for objects with color irregularities or varying edge states
Solution Approach 1:
The patent introduces a learning processor as an intermediary component between the image input unit and the determination unit. This learning processor automatically learns optimal features and thresholds from training images, acting as a mediator that transforms raw image data into accurate determination results without requiring complex manual feature extraction and threshold setting by users.
Solution Approach 2:
The learning processor performs self-learning and self-optimization by automatically extracting features and determining thresholds from training data. This self-service capability eliminates the need for users to manually configure complex inspection parameters, thereby improving determination accuracy while maintaining system simplicity.
2Ease of operation
If a learning processor is used to automatically learn features and thresholds, then user operation is simplified and determination accuracy improves, but erroneous input of training images may occur reducing learning effectiveness
Solution Approach 1:
The patent implements a feedback mechanism where the determination unit provides determination results back to the learning processor. This feedback loop allows the system to continuously improve by learning from both correct and incorrect determinations, enabling users to correct erroneous inputs and refine the learning model over time, thereby maintaining high reliability while keeping operation simple.
Solution Approach 2:
The system performs preliminary learning by automatically learning features and thresholds from training images before actual inspection begins. This preliminary action simplifies user operation during the inspection phase, while the reliability of learning data is ensured through the subsequent feedback mechanism that allows correction of any erroneous training inputs.
3Measurement precision
If multiple non-defective and defective product images are input for learning, then the classifier becomes more accurate, but the time required for learning and setup increases
Solution Approach 1:
The learning processor performs feature extraction and threshold determination in advance during the learning phase using multiple training images. This preliminary action prepares the classifier ahead of time, ensuring high classification accuracy during actual inspection while minimizing the time impact on production, as the learning process occurs separately from the inspection workflow.
Data Source
AI summary
To suppress erroneous input in inputting a non-defective product image and a defective product image, thereby increasing accuracy of distinguishing between a non-defective product image and a defective product image. An additional image that is added with an attribute as either one of a non-defective product and a defective product by a user is plotted in a feature space, and the probability that the attribute of the additional image is wrong is estimated. In the case in which the additional image is expected to have a wrong attribute, this result is notified. Result of selection whether to correct the attribute of the additional image by a user who receives the notification is received. A classifier generator 22 determines the attribute of the additional image on the basis of the result of selection and corrects a classification boundary in accordance with the determined attribute.