Using Transfer Learning in X-ray Void Detection
Ensuring the quality of electronic products often depends on the reliable detection of voids—small air pockets or gaps—in solder joints and other critical areas. Traditionally, void detection in X-ray images has relied on algorithms that require significant expertise and manual tuning, making them less accessible and time-consuming to set up. Operators often need to understand complex computer vision concepts, and even then, traditional methods can struggle with artifacts, intensity shifts, or irregular void shapes.
Nordson Test & Inspection has recently released a white paper that introduces a new approach: using deep learning and transfer learning to automate and simplify void detection in X-ray images. Transfer learning allows AI models to be pre-trained on large, relevant datasets and then quickly adapted to new products or inspection tasks with minimal additional data. This reduces the need for extensive manual labeling and parameter adjustment, making the inspection process more intuitive and efficient for operators.
The Nordson Test & Inspection team developed and tested this approach in collaboration with Harman, a leading electronics manufacturer. Their method involves labeling a small number of sample images, fine-tuning a pre-trained model, and then deploying it for inspection. The process is fast—training typically takes only a few minutes—and the resulting models can be reused across different product batches and even different products, thanks to the flexibility of transfer learning.
In their experiments, the team focused on two main inspection targets: BGA (Ball Grid Array) balls and solder joints. They trained models on one batch of samples and tested them on others, finding that the models maintained high accuracy (typically between 0.93 and 0.98) even without retraining. This demonstrates that a single model can be shared across multiple products, significantly reducing the effort required when new products are introduced.
The paper also discusses the importance of subjective labeling—different annotators may have different standards for what constitutes a void. By collaborating closely with industrial partners and standardizing labeling practices, the team ensured that their models met real-world requirements.
Overall, the results show that deep learning with transfer learning is not only effective for void detection but also highly practical for industrial environments. The approach minimizes setup time, reduces the need for expert intervention, and allows for ongoing improvement and adaptation as more data becomes available. This makes it a scalable and future-proof solution for manufacturers facing evolving inspection needs.
Nordson recently launched N-Intelligence, its Deep Learning and Advanced AI software solution for both MXI (Manual X-ray Inspection) and AXI (Automatic X-ray Inspection) systems. By combining the expertise of human inspectors with the consistency and speed of automated inspection, N-Intelligence enables faster, more accurate, and more reliable defect detection.
For more details and access to the full white paper, scan the QR-code. For direct inquiries, contact your regional Nordson Test & Inspection representative Partnertec BV.