ParsaVictor/pcb-solder-defect-detection-yolo26

PCB Solder & Assembly Defect Detection (YOLO26m)

A YOLO26-based object detector that finds soldering and assembly defects on populated PCBs: solder bridges, insufficient/excess solder, voids, missing or shifted components, burns, physical damage, and oxidation.

Part of the PCB Inspection Suite. Trained on ParsaVictor/pcb-solder-defect-detection.

Metrics

Split	mAP50	mAP50-95
val	0.905	0.647
test	0.895	0.657

Trained: YOLO26m, imgsz 800, 150 epochs (best at epoch 143), AdamW with a tuned low learning rate, extra augmentation (noise/motion-blur/JPEG) and an automatic rollback guard against training spikes.

Known weaknesses

• solder_void is frequently confused with insufficient_solder (~71% of void mistakes land there) — review both classes together on any flagged board
• burn recall is lower (~62%) — the class has the fewest training examples (513 instances)
• Performance drops on very dark/low-light photos from one of the source subsets

Per-class confidence thresholds (F1-optimal on val)

Class	Threshold
insufficient_solder	0.297
excess_solder	0.349
solder_bridge	0.417
solder_void	0.249
missing_component	0.381
shifted_component	0.389
burn	0.224
physical_damage	0.214
oxidation	0.228

A global fallback threshold of 0.228 is used for any class with too few validation examples to tune individually.

Usage

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

weights = hf_hub_download("ParsaVictor/pcb-solder-defect-detection-yolo26", "best.pt")
model = YOLO(weights)

results = model.predict("your_board_photo.jpg", conf=0.228)
results[0].show()

For best accuracy, apply the per-class thresholds above instead of one global confidence value — see thresholds.json in this repo.

Related

• Training dataset: pcb-solder-defect-detection
• Sibling model: pcb-bare-board-defect-detection-yolo26
• Full suite overview: pcb-inspection-suite

License

Research/development use — see the dataset license. Contact the author before commercial use.