Welding Inspection with Image Processing

Abstract

We studied a welding inspection for terminal welding of Fuel Pump with image processing. Color Extraction Method with blob analysis had gray zone of 13%. The gray zone is a vague area which is mixture area of good pieces and defective ones. There are two methods to determine a decision boundary. One is the Mahalanobis distance method that calculates the decision boundary given the distribution of learning data statistically. The other is SVM method that calculates the boundary that maximizes the margin of the border to identify all of the training data provided. We decided to adopt the SVM method in order to ensure the margin which does not flow out absolutely defective. First, we made a SVM tool for selecting the three features and could display the SVM border surface in 3D. But, the SVM method possibilities of misjudgment still exist for the defective work that cannot be covered by the supervised learning. Then, we applied SVM with the process capability index (CP). As a result, we could decrease excessive judgments as defective goods greatly.