Enhancing Image Segmentation Model in Computing Void Percentage With Mask RCNN

Abstract Automating quality control has been an ongoing effort, especially when manufacturing large flip-chips. One such method is through employing the convolutional neural network (CNN). This work studied the optimum setup of the Mask region-based convolutional neural network (RCNN) model’s accuracy in analyzing asymmetrical, large-amount ball grid array (BGA) flip-chip underfilling void formation and its size relative to the underfill region. Experimental images of the through-scan acoustic microscope (TSAM) of BGA underfill are collected, preprocessed, and trained with the Mask RCNN model by tweaking its backbone architecture and hyperparameter. Extraction of the detected region size is computed with a histogram. Otsu’s thresholding method and the model’s performance in generating the results with its accuracy relative to real-scale images are evaluated based on the customization done with the CNN and thresholding model. The Mask RCNN-ResNet101-FPN-Custom with Otsu’s thresholding method yields the best-performing result in both capturing void(s) in TSAM images up to 96.40% in accuracy and computing the void percentage relative to the underfilling region with a low percentage error of 1.70%. The study provides insight to improve further capturing and computing void presence and size, allowing manufacturers to leverage optimized CNN architecture and image segmentation thresholding algorithm to expedite automated quality checking in a manufacturing process, reducing lead cost.