Comparison of mask R-CNN and YOLOv8-seg for improved monitoring of the PCB surface during laser cleaning

Potting compounds and coatings protect electronic components in harsh environments, requiring careful removal for recycling or repair. This study introduces the innovative use of YOLOv8-seg and Mask R-CNN to enhance the precision and efficiency of the laser cleaning process for PCBs (Printed Circuit Boards). These models are utilized for two primary tasks: real-time segmentation for laser cleaning guidance and post-cleaning surface quality assessment. Real-time segmentation adapts cleaning strategies based on PCB surface states such as 'Bare-Cu', 'Complete-Removal', 'Incomplete-Removal', etc. Quality assessment ensures high-quality, damage-free surfaces post-cleaning. Both models were trained on an augmented dataset to improve robustness. In the initial test dataset, YOLOv8-seg (l), known for its speed, achieved an mAP50 (seg) of 82.8% at 3.98 FPS, proving suitable for time-sensitive laser cleaning processes due to its speed and precision. Mask R-CNN (ResNet-50) reached an mAP50 (seg) of 84.097% at 1.52 FPS, fulfilling real-time requirements with high precision. Although their visualization segmentation results on the initial test dataset vary, both models successfully address the previously mentioned tasks. When tested on a new dataset with unseen patterns it was shown that YOLOv8-seg excels at generalizing to new patterns while Mask R-CNN performs less effectively. This study confirms YOLOv8-seg's effectiveness in real-time PCB monitoring during laser cleaning, boosting automation and efficiency in PCB recycling.