Abstract Void inspection is crucial for insulated gate bipolar transistors. X-ray computed laminography (CL) is an imaging method for plate-like objects such as insulated gate bipolar transistors. Hence, voids in CL images should be identified automatically and accurately. This paper proposes a fully automatic method of void inspection in the welding layer between a direct-bonded ceramic and a baseplate. No-reference image sharpness assessment is used to find the welding-layer slice, which confirms that the error is within one layer as compared to well-trained operators. Next, semantic segmentation network DeepLabv3+ is modified to segment the solder region and voids. Compared with conventional segmentation algorithms based on the edge and threshold, the proposed method for segmenting the solder region is free of tuning parameters. Its global accuracy and the class average accuracy can be greater than 96%. In addition, for the segmentation of voids, the proposed algorithm is found to be faster than the region-growing algorithm with an improvement in speed of about hundreds of times. Our algorithm potentially enables objective assessment of the heat exchange efficiency of IGBTs in industry.

中文翻译：

---

基于计算机层析成像的绝缘栅双极晶体管缺陷自动检测

摘要 空洞检测对于绝缘栅双极晶体管至关重要。X 射线计算机层析成像 (CL) 是一种用于板状物体（例如绝缘栅双极晶体管）的成像方法。因此，应该自动准确地识别 CL 图像中的空洞。本文提出了一种全自动检测直接结合陶瓷与基板之间焊层空隙的方法。无参考图像清晰度评估用于找到焊接层切片，与训练有素的操作员相比，这证实了误差在一层内。接下来，修改语义分割网络 DeepLabv3+ 以分割焊料区域和空隙。与传统的基于边缘和阈值的分割算法相比，所提出的焊料区域分割方法无需调整参数。其全局准确率和类平均准确率可大于96%。此外，对于空隙的分割，发现该算法比区域增长算法更快，速度提高了约数百倍。我们的算法有可能实现对工业中 IGBT 热交换效率的客观评估。