CNN Preprocessing Filters for Versatile Defect Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing defect inspection systems face challenges in versatility due to varying inspection conditions across production lines, leading to overlooked defects or excessive inspections when using pre-learned identifiers, which are inefficient to adapt to different workpiece types and photographing conditions.
Innovation Solution
A defect inspection apparatus utilizing a convolutional neural network (CNN) with a preprocessing filter that adjusts input images based on learned models, allowing flexible utilization across diverse inspection objects and backgrounds, preventing misrecognition and false detections by generating filters tailored to specific defects and backgrounds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a pre-learned identifier is commonly used in various inspection conditions, then device complexity is reduced, but measurement precision deteriorates as defects may be overlooked due to different inspection conditions
Solution Approach 1:
The patent transforms the inspection image by changing parameters such as brightness, contrast, and color balance to make defect portions stand out more clearly. This parameter transformation allows a pre-learned identifier to maintain high detection accuracy across different inspection conditions without requiring relearning, thus resolving the contradiction between device complexity and measurement precision
Solution Approach 2:
The patent introduces an image transformation unit as an intermediary between the image acquisition unit and the identifier. This intermediary processes the input image to enhance defect visibility before it reaches the pre-learned identifier, enabling the identifier to maintain high precision without being exposed to varying inspection conditions directly
2Measurement precision
If the identifier is relearned according to each production line, then measurement precision is improved, but productivity deteriorates due to time required for reconstructing the identifier
Solution Approach 1:
The patent performs image parameter transformation in advance before the identifier makes its determination. By pre-processing the image to enhance defect visibility and standardize appearance, the system eliminates the need for relearning when switching production lines, thus maintaining both high precision and productivity
Solution Approach 2:
The system changes image parameters such as brightness, contrast, and color to create a standardized representation of the inspection object. This parameter standardization allows the pre-learned identifier to maintain high accuracy across different production lines without requiring time-consuming relearning processes
3Measurement precision
If image transformation is applied to enhance defect visibility, then measurement precision is improved, but device complexity increases due to additional processing steps
Solution Approach 1:
The patent applies parameter changes (brightness, contrast, color balance) to the input image to enhance defect visibility. These transformations are computationally efficient and can be implemented with minimal additional processing complexity while significantly improving defect detection accuracy
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The disclosure includes: a memory section (112, 130), storing a learned model and an internal parameter set in the learned model; an acquisition section (114), acquiring an image of an inspection object photographed under predetermined conditions; a preprocessing section (154), generating a predetermined preprocessing filter (136) according to a feature value in a preprocessing image being an image of the inspection object that is acquired by the acquisition section (114) and that includes a defect, and a feature value corresponding to the internal parameter, and generating a preprocessed image by applying the generated preprocessing filter (136) to an inspection image being an image of the inspection object that is acquired by the acquisition section (114) and converting the inspection image; and an inspection section (156, 170), inspecting the preprocessed image for presence or absence of the defect of the inspection object by using the stored learned model.