BondSight: wire-bond defect detector

BondSight finds defects in photos of wire-bonded chips. It draws a box around each problem, names it and gives a confidence. This is the trained model; the inspection station that runs it, the training notebooks and all the documentation are on GitHub: https://github.com/Avishek2410/bondsight

What it detects

Eight defect types: wire_sweep, sagging_loop, lifted_ball, lifted_stitch, off_center_ball, missing_wire, broken_wire and wire_short. Each type has its own confidence level, stored in the .json next to the model.

Files

File	What it is
wirebond_dfine_small_m3k.onnx	The detector: D-FINE Small (10.2 M parameters), fine-tuned for wire bonds. Input 640 × 640 RGB, outputs 300 candidate boxes with a score for each type.
wirebond_dfine_small_m3k.json	Class names, the confidence level for each type, and the tiling settings for large photos. Keep it next to the .onnx.
anomaly_backbone.onnx + .json	Optional: the feature extractor for BondSight's unusual-spot checker (ResNet18, ImageNet weights, unchanged).
CREDITS.md	Who made the photos and networks the model learned from. Keep it with the model files if you share them.

How well it works

Measured on test photos the model never saw during training:

Test set	Result
Real defects found (NTUST wire-bonding photos)	63%
Missing wires found (WHU close-ups)	82%
Bent wires found (Fuji power-module photos)	100%
Complete good boards wrongly flagged (hawk-team)	0%
Good chips wrongly flagged (lyra)	18%
Good modules wrongly flagged (Fuji Pass)	49%

Some of these test photos come from the same sources as the training data, so on a new product expect lower numbers until the model has been fine-tuned on that product's photos. It is a research prototype, not a validated inspection system. The full results, and how each earlier model did, are in docs/results.md.

Speed, CPU only with ONNX Runtime: 3.1 s for one 2048-pixel photo on a Raspberry Pi 5, 2.3 s on a laptop CPU.

How to use it

The simplest way is the BondSight inspector, which handles tiling of large photos, the per-type confidence levels, a web page for operators and a database:

git clone https://github.com/Avishek2410/bondsight
cd bondsight
python -m venv inspector/.venv
inspector/.venv/Scripts/python -m pip install -e synth -e inspector     # Linux/macOS: .venv/bin/

Then, in Python:

from huggingface_hub import hf_hub_download
from PIL import Image
from wirebond_inspector.detector import OnnxDetector

model = hf_hub_download("Avishek2410/bondsight", "wirebond_dfine_small_m3k.onnx")
hf_hub_download("Avishek2410/bondsight", "wirebond_dfine_small_m3k.json")   # settings, same folder

detector = OnnxDetector(model)
for f in detector.detect(Image.open("my_photo.jpg"), threshold=0.3):
    print(f.cls, round(f.score, 2), f.box)       # box = x0, y0, x1, y1 in pixels

Large photos are shrunk to half size and cut into overlapping 640 × 640 tiles automatically. How the detector, the tiling and the confidence levels work is explained, with the maths, in docs/technical.md.

Training

Fine-tuned from D-FINE Small COCO weights for 12 epochs on a Kaggle T4 GPU (AdamW, learning rate 2e-4, one-cycle schedule); the epoch with the lowest validation loss (epoch 6) was kept. Each epoch saw 6,367 pictures: drawings from BondSight's own picture generator and tiles from six public photo sets, including photos of good boards so the model learns what not to flag.

Credits and licence

The model files are released under CC BY 4.0: you can use and adapt them, including commercially, as long as you give credit to BondSight (Avishek Mondal) and keep CREDITS.md with them.

The model learned from CC BY 4.0 photos by National Taiwan University of Science and Technology, WHUrobotics, hawk-team, lyra-wikbf and pg-87jvm (Roboflow Universe), and starts from D-FINE (Peng et al., 2024; Apache-2.0). The checker uses torchvision's ResNet18 (BSD-3-Clause). Details of who made what and what was changed are in CREDITS.md.

Citation

@software{mondal_bondsight_2026,
  author  = {Mondal, Avishek},
  title   = {BondSight: open-source AOI for wire bonds},
  year    = {2026},
  version = {0.1.0},
  url     = {https://github.com/Avishek2410/bondsight}
}