Cascaded Approach to Defect Location and Classification in Microelectronic Bonded Joints: Improved Level Set and Random Forest

The monitoring of bonded joints in microelectronics packaging is generally done manually and offline. However, this method is inefficient and of low precision due to limited personal sensing and experience. This article proposes a hybrid cascade approach with an improved level set and a random forest to locate and automatically classify defective joints in microelectronics packaging. A grayscale variance-based pixel neighborhood is introduced to accurately locate the joint, and an improved gray projection is used to remove the redundant nonjoint area. We have used an improved level set algorithm to segment joint defects and extract their dominant features using KPCA. Finally, a random forest is used to classify the features extracted by KPCA and determine the defect categories. The results have indicated that the grayscale variance-based pixel neighborhood could effectively locate the joint and that KPCA could identify effective joint features. The accuracy of the random forest classification has reached 0.91, offering a novel solution for joint quality monitoring in microelectronics manufacturing.