PCBInspect-BodyDefect

Part of the SentinelPCB defect-inspection router: a region classifier dispatches each component ROI crop to a region-specific defect classifier. Sibling repos: PCBInspect-Region, PCBInspect-BodyDefect, PCBInspect-LeadDefect, PCBInspect-TextDefect.

Companion structural-feature detector (unrelated task — detects MountingHole/ComponentBody/ SolderJoint/Lead, not defects): PCBInspect-AI.

Role

Defect classifier for crops routed as Body by the region classifier. The hardest of the three specialists — six classes with real class imbalance.

Model

• Base: yolo26s-cls (Ultralytics), classification head, fine-tuned on AOI component-ROI crops.
• Export: ONNX, opset 17, no NMS (classification only) — single input images (1, 3, 640, 640) RGB, normalized /255, NCHW; single output output0 (1, 6) raw logits (apply softmax yourself for probabilities).
• Classes (6), index order = labels.json: ForeignMaterial, Golden, MissingPart, Shift, Tombstone, WrongPart.

Data

Trained on a proprietary AOI dataset of SMT component-ROI crops (paired defect-free reference + defective capture per physical site), not publicly released. Split is grouped by physical capture site (never by raw image) so a component's reference and defect crop never straddle train/val/test.

Metrics

val (top-1 0.948, macro-F1 0.875, n=784):

class	precision	recall	f1	support
ForeignMaterial	0.996	0.975	0.985	276
Golden	1.000	1.000	1.000	200
MissingPart	0.706	0.706	0.706	17
Shift	0.929	0.897	0.913	146
Tombstone	0.773	0.739	0.756	23
WrongPart	0.851	0.934	0.891	122

test (top-1 0.941, macro-F1 0.871, n=780):

class	precision	recall	f1	support
ForeignMaterial	0.996	0.963	0.979	267
Golden	0.995	1.000	0.998	200
MissingPart	1.000	0.500	0.667	22
Shift	0.901	0.938	0.919	146
Tombstone	0.833	0.769	0.800	26
WrongPart	0.813	0.916	0.862	119

Limitations

MissingPart and Tombstone are the weak classes (94 and 131 raw examples total, concentrated on 12 and 15 distinct part-numbers respectively). Test: MissingPart precision 1.00 / recall 0.50 (conservative — misses about half, never false-alarms); Tombstone F1 0.80. More examples across more board/package designs would improve both; see the companion data-exploration notebook.

Usage

import onnxruntime as ort
import numpy as np
from PIL import Image

sess = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
img = Image.open("crop.jpg").convert("RGB").resize((640, 640))
x = (np.asarray(img, dtype=np.float32) / 255.0).transpose(2, 0, 1)[None, ...]
(logits,) = sess.run(None, {"images": x})
probs = np.exp(logits) / np.exp(logits).sum()
print(probs)