PCB CT image element segmentation model based on boundary-attention-guided finetuning.

Area of Science:

• Computer Vision
• Non-destructive Testing
• Image Segmentation

Background:

• Computed Tomography (CT) is vital for non-destructive testing of Printed Circuit Boards (PCBs).
• Element segmentation in PCB CT images is challenging due to grayscale inconsistencies and layer penetration, leading to inaccurate boundaries.
• Existing pretraining and finetuning methods struggle with these image artifacts, impacting semantic understanding.

Purpose of the Study:

• To address the limitations in segmenting PCB CT image elements, specifically poor boundary definition.
• To enhance the accuracy and reliability of element segmentation in PCB CT analysis.

Main Methods:

• A novel boundary-attention-guided finetuning model (BAG-FTseg) is proposed for PCB CT image element segmentation.
• An improved boundary detection algorithm enhances the model's ability to sense element boundaries.
• An Attention Feature Fusion module integrates boundary features effectively using an attention mechanism for improved segmentation.

Main Results:

• The proposed BAG-FTseg model achieved an 89.5% mean Intersection over Union (mIoU) on the PCB CT dataset.
• The model demonstrated a 0.9% improvement in overall mIoU compared to the baseline model.
• Boundary-specific mIoU reached 69.5%, a significant 5.3% increase over the baseline, highlighting improved boundary segmentation.

Conclusions:

• The developed method enhances the accuracy of PCB element boundary segmentation.
• The attention mechanism improves feature fusion efficiency, contributing to practical advancements in PCB analysis.
• This approach offers significant practical value for non-destructive testing and quality control of PCBs.