講演情報
[11a-A13-4]Automatic Detection of Threading Dislocations in Synchrotron X-ray Topography
Images of 4H-SiC Using YOLOv8s and CBAM-Augmented Faster R-CNN
〇(M2)WEIYUAN XIA1, Rui Zhou1, Yuhui Huang1, Takayoshi Shimura1 (1.Waseda Univ.)
キーワード：
silicon carbide、X-ray topography、deep learning
Threading screw and edge dislocations (TSDs/TEDs) in 4H-SiC degrade high-power device performance, making wafer-scale inspection increasingly important. Synchrotron X-ray topography (XRT) provides non-destructive, high-resolution imaging, but analyzing >150-Mpixel whole-wafer composites is a throughput bottleneck. We develop and compare two deep object-detection frameworks—YOLOv8s and a CBAM-augmented Faster R-CNN (ResNet-50-FPN-v2 backbone)—to automatically detect and localize both dislocation types in tiled XRT images. On wafer-scale composites, Faster R-CNN markedly outperforms YOLOv8s for the weak, low-contrast TEDs, reaching 99.66% recall and 98.48% precision (vs. 91.98% and 91.82%), while raising TSD precision to 98.53%. Its residual backbone, multi-scale features, and attention-guided two-stage design preserve faint dislocation signals that one-stage detectors tend to lose, enabling accurate, fully automated wafer-scale dislocation mapping for next-generation SiC power devices.