AI 中文总结
本研究对比了基于规则、CNN及YOLOv8三种位错自动检测方法，发现YOLOv8经自适应高斯分块算法与调整置信度阈值优化后，在GaN的ECCI图像上实现了高精度与高速度，适用于半导体缺陷高通量表征。
AI 中文摘要
通过电子通道衬度成像（ECCI）量化半导体材料中的 threading 位错，面临手动分析速度缓慢的严重瓶颈。本研究以统计真值为基准，在氮化镓（GaN）的 ECCI 显微图像上对三种自动检测流程进行了基准测试。经典的基于规则的计算机视觉方法因需要大量逐图像调参而被证明不可靠；而基于卷积神经网络（CNN）的多阶段分类与定位方法达到了87%的准确率，但单张测试图像耗时较长，且在高密度区域效率较低。相比之下，统一的单阶段 You Only Look Once（YOLOv8）架构在多张图像的7451个位错上实现了98.6%的计数准确率，以及98.7%的精确率和98.7%的召回率，同时推理速度极快。YOLOv8的成功部署需解决两个特定领域的机器学习挑战：第一，为缓解模型的尺度敏感性，开发了自适应高斯分块尺寸算法以优化每张图像的视野；第二，部署的置信度阈值（0.025）与YOLOv8建议的推理默认值（0.25）存在显著差异，这将基准图像某一子区域的计数准确率从90.8%提升至99.3%，此处低置信度分数代表物理信号强度而非分类歧义。这种统一的尺度自适应方法展现了其在高通量、定量半导体缺陷表征中的实际可行性。
英文摘要
Quantifying threading dislocations in semiconductor materials via electron channelling contrast imaging (ECCI) is heavily bottlenecked by slow manual analysis. This work benchmarks three automated detection pipelines on ECCI micrographs of gallium nitride (GaN) against a statistical ground truth. A classical rule-based computer vision approach proved unreliable due to extensive per-image tuning requirements, while a convolutional neural network (CNN)-based multi-stage classification and locator method achieved 87% accuracy but required significant time for a test image and was less efficient in high-density regions. By contrast, a unified single-stage you only look once (YOLOv8) architecture achieved a counting accuracy of 98.6%, alongside 98.7% precision and 98.7% recall across 7451 dislocations over multiple images, with rapid inference times. Successful deployment of YOLOv8 required addressing two domain-specific machine learning challenges. First, to mitigate the model's scale sensitivity, an adaptive gaussian tile-sizing algorithm was developed to optimise the field of view per image. Second, the deployed confidence threshold (0.025) differed substantially from the suggested inference default value for YOLOv8 (0.25), raising counting accuracy from 90.8% to 99.3% on a subsection of the benchmark image. In this case, low confidence scores represented physical signal strength rather than classification ambiguity. This unified, scale-adaptive approach demonstrates practical viability for high-throughput, quantitative semiconductor defect characterisation.
Comments31 pages, 9 figures, 5 tables
