A deep learning based dislocation detection method for cylindrical silicon growth process.

Autor: Yuting, She, Hongxing, Li
Předmět:
Zdroj: Applied Intelligence; Apr2023, Vol. 53 Issue 8, p9188-9203, 16p
Abstrakt: In the process of preparing heavily doped silicon crystals for power electronics and solar power by the Czochralski(CZ) method, dislocations occur due to the presence of impurities and other factors. The presence of dislocations can affect the quality of single-crystal silicon crystalline columns. Therefore, we proposed the improved Yolov4-tiny model (Yolo-SPI) for detecting the occurrence of dislocations. For resolving the problem of low detection accuracy of the original model, we improve the neck part of the model by referring to the structure of the Path Aggregation Network (PAnet). Furthermore, we propose a feature enhancement module to improve the feature extraction ability of the model and introduce depthwise separable convolution to reduce the parameters. We produced the single-crystal silicon habit line dataset by using the industrial camera. The experimental results on our dataset show that the Yolo-SPI model outperforms Ghostnet-Yolov4, Mobilenetv3-Yolov4, EfficientDet-v0, and Nanodet. The Yolo-SPI model improves the Precision from 73.28% to 98.01% compared to the original Yolov4-tiny model and the Recall is also improved to 97.51%. At the same time, the number of parameters in the Yolo-SPI model is decreased from 5.87M to 2.05M compared to the Yolov4-tiny model. Our model also beats other models in the speed of detection, reaching 133FPS. In practical applications, our model achieves higher accuracy and faster detection speed. [ABSTRACT FROM AUTHOR]
Databáze: Complementary Index