RFchips, as core components of modern communication systems, play a critical role in determining the reliability and stability of circuit connections. The quality of the bond wire and bond ball in RFchips directly affects the performance of the entire system. Traditional automatic optical inspection methods have shown limitations in the identification of bond balls and bond wires in RFchips, including a high false positive rate and reliance on manual re-inspection. To address these issues, this research proposes an improved Mask R-CNN-based instance segmentation method for RFchips that achieves accurate and efficient identification of bonding structures.

Challenges in RFchip Bonding Detection

Bond balls and bond wires in RFchips are often small and diverse in shape, making it difficult to detect them accurately using traditional methods. This characteristic requires a more sophisticated detection model that can handle varied sizes and forms, especially in production environments where speed and accuracy are critical.

Key Improvements in the Mask R-CNN Model for RFchip Detection

In this study, the Mask R-CNN network structure was enhanced in two key ways to improve its performance in detecting bond balls and bond wires in RFchips:

Optimization of the Prior Box Generation Mechanism:
The original Mask R-CNN feature pyramid network (FPN) layers P2, P5, and P6 had an excessive number of prior boxes. This was optimized by reducing the number of prior boxes from 260,000 to 190,000, significantly increasing the model’s inference efficiency without sacrificing accuracy.

Innovative Data Augmentation via Collision Detection:
To address the issue of limited training data, the study introduced a data augmentation strategy based on collision detection. This strategy randomly transforms bond balls and bond wires and then places them into realistic RFchip backgrounds. This method effectively expanded the training dataset, reducing annotation costs and improving the robustness of the model.

Experimental Results and Performance

The model was trained using RFchip image data collected directly from production lines, which included a variety of bond ball and bond wire shapes and background types. The results showed that the improved Mask R-CNN model achieved a mean average precision (mAP) of 85.23% and an average intersection-over-union (IoU) of 71.27% for RFchip bond ball and bond wire segmentation tasks. The inference time per RFchip image was only 1.168 seconds, significantly outperforming traditional methods such as UNet and FCN.

Notably, the model excelled in segmenting the first and second bond point bond balls of RFchips, with an average precision (AP) of over 87%. This demonstrated the model’s superiority in identifying the intricate microstructures of RFchips.

Real-World Application and Future Work

Currently, this recognition method has been implemented and tested on production lines for RFchips. It has proven to be effective and practical in real industrial environments. The method shows strong potential in addressing real-world challenges in RFchip manufacturing, especially in the context of high-speed inspection.

Future work will focus on optimizing the segmentation performance for larger bond wires in RFchips, further enhancing the method’s applicability in complex detection tasks within the RFchip manufacturing process.