Reproduce “AOI-SSL: Self-Supervised Framework for Efficient Segmentation of Wire-bonded Semiconductors In Optical Inspection” on OpenResearch

About this paper

Joaquín Figueira, Rob Van Gastel, Giacomo D'Amicantonio, Zhuoran Liu, Ioan Gabriel Bucur, Faysal Boughorbel, Egor Bondarev

The AOI-SSL framework introduces a self-supervised approach for segmenting wire-bonded semiconductors in optical inspection, enabling efficient adaptation with limited labeled data. It achieves superior segmentation performance and faster deployment compared to supervised and generic self-supervised baselines, particularly for fine-grained structures.

Key insights

• Domain-specific self-supervised pre-training, especially MAE's pixel-level reconstruction objective, is highly effective for dense downstream tasks in low-semantic-diversity industrial domains, outperforming DINO and iBOT.
• Hybrid Vision Transformer architectures like FasterViT, which integrate convolutional elements, significantly improve the segmentation of fine-grained structures (e.g., wires) compared to pure ViT architectures.
• Retrieval-based segmentation provides an efficient, training-free adaptation mechanism that can achieve competitive performance with significantly less labeled data than fully supervised methods, offering high throughput and robust detail preservation in single-device scenarios.

Summary by alphaXiv.