Moszer777/pcb-aoi-yolo-rmutt

PCB AOI · YOLO26 electronic component detector (RMUTT)

YOLO26 models that find 11 kinds of electronic components on printed circuit boards. They are the vision part of a PCB automated optical inspection (AOI) machine built at Rajamangala University of Technology Thanyaburi (RMUTT): a 4K camera on an XY gantry visits every point of a board, the model detects the parts in each frame, and the machine compares them with a golden reference board to decide PASS / FAIL for every part (OK, MISSING, WRONG type, EXTRA).

ภาษาไทย: โมเดล YOLO26 สำหรับตรวจจับชิ้นส่วนอิเล็กทรอนิกส์ 11 ชนิดบนแผ่น PCB ใช้ในเครื่องตรวจจับความสมบูรณ์ของแผ่น PCB ด้วยปัญญาประดิษฐ์ ของมหาวิทยาลัยเทคโนโลยีราชมงคลธัญบุรี โมเดลหลักคือ new_dataset_v6_split_v1_aug_v2_m/weights/best.pt (mAP50 98.8% และ F1 98.1% บนชุดทดสอบ) ส่วนรุ่นล่าสุดคือ v7

Which file to use

Model	Path	Notes
Main model (v6)	new_dataset_v6_split_v1_aug_v2_m/weights/best.pt	Used on the machine and in the thesis. Evaluated on the held-out test set.
Latest (v7)	new_dataset_v7_split_v1_aug_v2_m/weights/best.pt	Same recipe on v7 data (v6 + the data_fix_con set, 40 more images). Validation results only so far.

Every other folder is an earlier experiment (25 training runs in total), kept with its args.yaml, results.csv and plots so the work can be traced. Each folder is one run: weights/best.pt, weights/last.pt, training curves, confusion matrices and sample batches.

Classes

buzzer · capacitor · clock · connector · display · ic · led · potentiometer · resistor · switch · transistor

(class ids 0–10 in this order). Both through-hole and surface-mount parts are labelled. The diode class used in earlier runs was dropped: it had too few examples to train or measure reliably.

Results

Main model (v6), held-out test set

43 images and 1,372 boxes that were never used for training or model selection, image size 640 px.

mAP@0.50	mAP@0.50:0.95	Precision	Recall	F1
98.8%	95.8%	97.2%	99.1%	98.1%

Per class (test set):

Class	Precision	Recall	AP50	AP50-95
buzzer	96.1	100.0	99.5	97.1
capacitor	97.5	98.8	98.7	96.0
clock	92.1	100.0	93.7	93.7
connector	99.3	97.5	99.4	96.7
display	97.8	98.6	99.5	97.9
ic	98.1	100.0	99.0	97.2
led	97.4	98.4	99.5	97.9
potentiometer	92.8	100.0	99.5	99.5
resistor	98.4	98.7	99.1	88.4
switch	99.3	99.6	99.5	98.7
transistor	99.9	98.7	99.4	91.0

potentiometer has only 2 boxes and clock 12 boxes in the test set, so their numbers move more than the others.

Validation (best epoch of each run)

Run	Data	mAP@0.50	mAP@0.50:0.95	Precision	Recall
v6 new_dataset_v6_split_v1_aug_v2_m	43 images	97.2%	92.7%	95.4%	96.6%
v7 new_dataset_v7_split_v1_aug_v2_m	47 images	97.3%	93.4%	96.6%	96.4%

The two validation sets differ by 4 images, so these rows are not a direct comparison.

Speed (main model, 640 px)

Median time per image (preprocess + inference + postprocess) on 114 real images from the machine:

Device	Backend	Time / image	Images / s
NVIDIA Jetson Orin Nano (on the machine)	CUDA	72.6 ms	13.4
Apple Silicon Mac	MPS	26.4 ms	38.1

Training

Base model	yolo26m.pt (pretrained, transfer learning)
Image size / batch	640 / 16
Epochs	60 (patience 20); v6 best at epoch 46
Optimizer	auto, lr0 0.01, lrf 0.01, momentum 0.937, weight decay 0.0005, warmup 3
Online augmentation	mosaic 1.0 (off for the last 10 epochs), fliplr 0.5, hsv 0.02 / 0.8 / 0.4, scale 0.5, translate 0.1
Hardware	Apple Silicon Mac, PyTorch MPS, AMP
Time	v6 2.2 h · v7 2.5 h

Data

Photos of real boards, labelled for this project and split before augmentation:

	v6 train	v6 valid	v6 test	v7 train	v7 valid	v7 test
Original images	344	43	43	376	47	47
Images used	1,032	43	43	1,128	47	47
Boxes	33,843	1,293	1,372	36,525	1,401	1,494

Offline augmentation, train split only, ×3 (the original plus 2 copies): horizontal and vertical flips, 90° / 180° / 270° rotations, brightness ±15 and exposure ±10%, seed 42. v7 is v6 merged with the data_fix_con set, each image kept in its original split; no exact or near-duplicate images were found across the two sources or across splits.

Usage

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download(
    "Moszer777/pcb-aoi-yolo-rmutt",
    "new_dataset_v6_split_v1_aug_v2_m/weights/best.pt",
)
model = YOLO(weights)

result = model.predict("board.jpg", imgsz=640, conf=0.25)[0]
for box in result.boxes:
    name = result.names[int(box.cls)]
    x1, y1, x2, y2 = box.xyxy[0].tolist()
    print(name, round(float(box.conf), 3), ((x1 + x2) / 2, (y1 + y2) / 2))

On the machine the detections are matched to a golden reference: each reference point takes the nearest unclaimed detection within 50 px; the same type is OK, a different type is WRONG, none is MISSING, and leftover detections are EXTRA. Any MISSING or WRONG (or EXTRA, when enabled) makes the point FAIL. The machine also takes several frames per point (default 5) and counts a part as present when it is found in at least 80% of them.

Limitations

• Trained on the boards and lighting of this project. Other boards, cameras or lighting need new data to reach the same accuracy.
• The scores above are box detection scores. Board-level PASS / FAIL accuracy depends on the reference profile and the matching settings and has to be measured on real good and faulty boards.
• No diode class.
• max_det is 300 per image (YOLO26 default); a single frame with more parts than that will miss some.

Links

• Project site: https://rmutt-pcb-aoi-landing.vercel.app
• Training dashboard (all runs): https://pcb-training-dashboard.vercel.app
• Source code: https://github.com/moszer/PCB_DETECT_RMUTT_GUI

Department of Electronics and Telecommunication Engineering, Faculty of Engineering, Rajamangala University of Technology Thanyaburi (RMUTT), 2026.