Optical plasma boundary detection using improved YOLOv8 segmentation network on EAST tokamak

Abstract

The stable optical detection of plasma boundary is expected to play a vital role in the shape control of long-pulse plasma in future tokamaks. However, optical plasma boundary detection under various illumination conditions and complex discharge conditions is still quite challenging. In this study, an optical plasma boundary detector based on an improved YOLOv8 segmentation algorithm is proposed. Through the Retinex image enhancement algorithm, it is found that the optical boundary information is contained in the image data obtained under different illumination conditions, and the optical boundary is labeled to establish the optical plasma boundary dataset. After training, the YOLOv8n-seg detector achieves a mean pixel accuracy (MPA) of 0.872 and mean intersection over union (mIoU) of 0.74 on the test set. To improve the overall contour detection capability of the detector for optical plasma boundary, the YOLOv8n-seg detector is enhanced by adding the P6 feature layer, and the YOLOv8n-seg-P6 detector exhibits an enhanced MPA of 0.877 and mIoU of 0.75. After introducing the convolutional block attention module (CBAM), the MPA and mIoU of the YOLOv8n-seg-P6-CBAM detector are further improved to 0.901 and 0.79, respectively. The YOLOv8n-seg-P6-CBAM detector can accurately detect optical plasma boundaries under different illumination performance and different configurations.