ParsaVictor/pcb-component-detection-yolo26

PCB Component Detection (YOLO26s, two-stage)

Two YOLO26s detectors that find and name electronic components on full-resolution PCB photos, plus two Random-Forest refiners for explainable type correction.

Part of the PCB Inspection Suite. Trained on ParsaVictor/pcb-component-detection (458 boards, 108k annotated components, tiled at native resolution — not downscaled to 640px).

Why two stages

Board photos are 2000-4900px wide. Downscaled to 640px, a 0402 resistor is ~6px and undetectable. Tiling at native resolution (1024px tiles) gives ~19px parts instead. Finding where a component is turns out to be far easier than naming it, so:

• any_best.pt — 1-class locator ("component"): finds every element on the board, whatever it is.
• fine_best.pt — 13-class typer: names the type, initialised from the locator's weights.
• tree_rf.joblib / tree_rf_boards.joblib — Random-Forest classifiers (33 hand-engineered CV features) that refine uncertain type predictions in an explainable way (93% holdout accuracy on isolated photos, 89% at p≥0.5 on training-board crops).

Metrics

Model	Split	P	R	mAP50	mAP50-95
any (locator)	val (best epoch)	0.773	0.771	0.809	0.427
any (locator)	test (46 unseen boards)	0.726	0.693	0.732	0.376
fine (13-class typer)	test	0.488	0.470	0.432	0.263

Per-class (fine, test)

Class	P	R	mAP50
IC	0.792	0.871	0.826
Electrolytic Capacitor	0.903	0.618	0.744
LED	0.887	0.477	0.652
Connector	0.502	0.665	0.538
Switch/Button	0.158	0.545	0.520
Capacitor	0.523	0.610	0.504
Resistor	0.499	0.467	0.426
Marker	0.597	0.312	0.381
Transistor	0.355	0.488	0.324
Oscillator	0.431	0.381	0.299
Other	0.288	0.284	0.204
Inductor	0.226	0.201	0.125
Diode	0.182	0.189	0.067

Trained: YOLO26s, imgsz 1024, 80 epochs, Tesla T4.

Usage

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

any_w = hf_hub_download("ParsaVictor/pcb-component-detection-yolo26", "any_best.pt")
fine_w = hf_hub_download("ParsaVictor/pcb-component-detection-yolo26", "fine_best.pt")

locator = YOLO(any_w)   # every element, type unknown
typer = YOLO(fine_w)    # 13-class type detector

results = typer.predict("your_board_photo.jpg", conf=0.20)
results[0].show()

Board photos are large — tile at 1024px (matching training) rather than resizing the whole image for best results. See the GitHub repo for the full two-stage + Random-Forest refinement pipeline (predict.py).

Related

• Training dataset: pcb-component-detection
• Full pipeline + notebooks: GitHub: pcb-component-detection-yolo26
• Sibling models: pcb-bare-board-defect-detection-yolo26, pcb-solder-defect-detection-yolo26
• Full suite overview: pcb-inspection-suite

License

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