Surface Defect Detection Using Consecutive Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated quality assurance methods for surface modification techniques, such as laser-beam soldering and welding, face challenges in achieving high accuracy and efficiency due to the need for extensive training datasets and the risk of overfitting, leading to low detection accuracy and high false-negative rates for surface defects.
Innovation Solution
A computer-implemented method using a neural network to classify image sequences of surface regions, where individual images are assigned to defective or non-defective classes, and a defect signal is output if a predetermined number of consecutive images are classified as defective, leveraging pre-trained networks like ResNet50 and additional machine learning techniques for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network is trained with a large number of training datasets to achieve high detection accuracy, then the detection accuracy improves, but the time and resources required for generating training datasets increase significantly
Solution Approach 1:
The patent applies transfer learning by using a neural network that has been pre-trained on a large dataset (ImageNet with 1.4 million images) before fine-tuning it on the specific defect detection task. This preliminary training on a general dataset provides the network with robust feature extraction capabilities, reducing the need for extensive task-specific training data and significantly decreasing the time required to generate training datasets for the specific application.
2Measurement precision
If manual visual inspection is used to ensure quality control, then the accuracy of defect detection is maintained, but the labor intensity and time consumption increase
Solution Approach 1:
The patent replaces the mechanical system of manual visual inspection with an automated optical inspection system using a neural network. The system captures images of the surface region during or after the surface modification process and uses the trained neural network to automatically detect defects, thereby eliminating labor-intensive manual inspection while maintaining or improving detection accuracy and significantly increasing productivity.
3Ease of operation
If image recording from the side facing away from the laser is used for quality assurance, then the measurement process is simplified, but the accuracy decreases for surfaces with extremely high quality
Solution Approach 1:
Instead of recording images from the side facing away from the laser (as in conventional methods), the patent positions the camera to record images from the side facing the laser. This inverted approach allows direct observation of the surface region during the surface modification process, enabling accurate detection of defects on high-quality surfaces while maintaining operational simplicity through real-time monitoring.
4Speed
If a high frame rate is used to capture individual images for defect detection, then the real-time detection capability improves, but the computational load and processing time increase
Solution Approach 1:
The patent extracts and processes only the most relevant features from the captured images using the trained neural network, rather than analyzing all image data in full detail. The network has learned to identify critical defect patterns during training, allowing it to rapidly classify images at high frame rates with reduced computational load by focusing only on the most discriminative features necessary for defect detection.
Data Source
AI summary
A method for determining defects that occur while carrying out a surface modification method of a surface region of a component includes providing an image sequence of a surface region to be assessed. Each image of the image sequence shows an image detail of the surface region and the image details of the individual images overlapping at least partially. The method further includes assigning the individual images to at least two image classes, where at least one image class among the at least two image classes is indicative of a defective image class. The method further includes checking whether a set of individual images of a predeterminable number of directly consecutive individual images in the image sequence have been assigned to the defective image class, and outputting a defect signal.