Location and timestamp-based chip contour detection using LWMG-YOLOv5

Abstract

In the fab, semiconductor manufacturers often use deep learning approaches for chip contour detection to shorten automated optical inspection to minimize the loss of production costs and lower power consumption in chip contour detection for realizing energy-efficient computing. However, YOLOv5 and GSEH-YOLOv5 models have sacrificed their accuracy to improve the operational speed. MobileNetv3-YOLOv5 model can enhance the accuracy but lacks high-speed operation. Therefore, this study presents a light version of MobileNetv3-YOLOv5 model with ghost convolution, abbreviated LWMG-YOLOv5, to speed up chip contour detection because this architecture can reduce the number of model parameters and computational burden at the same time. As a result, the proposed approach can outperform the other methods by getting a 3.62% speed-up in chip contour detection to gain a better manufacturing advantage in increasing the chip yields by 1.7% and reducing the loss of production costs by 1.83% significantly.