yolov8-industrial-finetune

name

yolov8-industrial-finetune

description

Use this skill for defect detection and instance segmentation on industrial products — magnets, metal parts, PCBs, ceramics, plastics, or any manufactured component with surface defects. The default base model is YOLOv8n-seg. When input images are single-channel (grayscale) TIF files — common with industrial cameras and microscopes — this skill applies the correct first-layer weight adaptation: average the 3-channel pretrained weights across the channel dimension and use that to initialize a 1-channel input model, then fine-tune. This is better than training from scratch or naively converting the image to RGB. Trigger this skill when the user mentions: "defect detection", "surface inspection", "magnet", "metal part", "PCB inspection", "TIF images", "single channel", "grayscale industrial", "fine-tune YOLOv8", "1-channel input", "instance segmentation defects", "train on my dataset". Always prefer this over generic training advice for industrial inspection tasks.

YOLOv8n-seg Industrial Defect Fine-Tuning

Opinionated, production-tested training pipeline for industrial defect detection and segmentation. Covers the full path from raw TIF dataset → trained model → deployment.

---

Default Model Choice: YOLOv8n-seg

For simple industrial products (magnets, stamped metal, cast parts, PCBs):

Factor	Why YOLOv8n-seg
Speed	Nano = fastest inference, fits Jetson/edge
Accuracy	Sufficient for surface defects with good data
Segmentation	Mask output enables area measurement, not just bbox
Training data	Converges well with 200–2000 labeled images
Upgrade path	Easy to swap to YOLOv8s-seg / m-seg if accuracy insufficient

When to upgrade from nano:

• mAP plateau below acceptable threshold after augmentation → try yolov8s-seg
• Defects are very small (< 8×8 px on 640 input) → try yolov8m-seg with SAHI
• Multiple complex defect classes (> 8 classes) → try yolov8m-seg

---

Phase 1: Input Image Check

Before anything else, determine the image type:

from PIL import Image
img = Image.open("sample.tif")
print(img.mode, img.size)  # e.g. "L 2048x2048" = grayscale, "RGB 1920x1080" = color

Result	Action
L or I;16 or I (1-channel grayscale)	→ Use grayscale fine-tune path (Phase 2A)
RGB or RGBA (3-channel color)	→ Use standard fine-tune path (Phase 2B)
16-bit (I;16)	→ Normalize to 8-bit first (see Phase 2A note)

---

Phase 2A: Grayscale Input — First-Layer Weight Averaging

This is the key technique for 1-channel TIF inputs.

Why this works better than alternatives

Approach	Problem
Convert gray→RGB (replicate channel 3×)	Works but wastes memory and compute — 3× the input data
Train from scratch with in_channels=1	Loses all ImageNet pretrained features — needs much more data
Average pretrained first-layer weights ✅	Preserves all learned feature detectors, correct channel count, fast convergence

The Math

YOLOv8's first conv layer has weights of shape [64, 3, 3, 3] (64 filters, 3 input channels, 3×3 kernel). Average across the channel dimension → [64, 1, 3, 3]. This maps each grayscale pixel to the same feature space the model expects.

Implementation

Read references/grayscale-finetune.md for the complete ready-to-run script.

Key steps the script performs:

1. Load yolov8n-seg.pt pretrained weights
2. Rebuild model with ch=1 (1 input channel)
3. Extract first conv weights: w = model.model[0].conv.weight.data (shape [64, 3, 3, 3])
4. Average: w_gray = w.mean(dim=1, keepdim=True) (shape [64, 1, 3, 3])
5. Assign: model.model[0].conv.weight.data = w_gray
6. Save modified model → use as starting checkpoint for training
7. Train with model.train(data=..., epochs=..., imgsz=640)

TIF 16-bit Normalization

If images are 16-bit (common with industrial cameras):

import cv2, numpy as np
img16 = cv2.imread("sample.tif", cv2.IMREAD_UNCHANGED)  # uint16
img8  = cv2.normalize(img16, None, 0, 255, cv2.NORM_MINMAX).astype(np.uint8)
cv2.imwrite("sample_8bit.tif", img8)

Do this as a preprocessing step for the whole dataset before training.

---

Phase 2B: Standard RGB Fine-Tune

For RGB images, standard Ultralytics fine-tune:

from ultralytics import YOLO
model = YOLO("yolov8n-seg.pt")
model.train(
    data    = "data.yaml",
    epochs  = 150,
    imgsz   = 640,
    batch   = 16,
    lr0     = 0.01,
    lrf     = 0.01,
    warmup_epochs = 3,
    device  = 0,
    project = "runs/defect",
    name    = "yolov8n_seg_v1",
)

---

Phase 3: Training Configuration

Read references/training-config.md for full hyperparameter guidance, augmentation settings, and dataset YAML templates.

Key hyperparameters for industrial defects

Parameter	Default	Industrial recommendation	Reason
imgsz	640	640 or 1280	Use 1280 if defects are small (< 20px)
epochs	100	150–300	Industrial datasets are usually small — need more epochs
batch	16	8–16 (GPU VRAM dependent)
lr0	0.01	0.005–0.01	Lower for small datasets to avoid overfitting
lrf	0.01	0.001	Cosine decay to very low LR
mosaic	1.0	0.5–1.0	Reduce if defects are positionally critical
degrees	0.0	0–180	Enable if defects are rotation-invariant
fliplr	0.5	0.5	Usually safe
flipud	0.0	0.5	Enable if top/bottom symmetric
hsv_s	0.7	0.2	Reduce — industrial images have controlled lighting
hsv_v	0.4	0.3
copy_paste	0.0	0.3–0.5	Very effective for defect segmentation with few samples
mixup	0.0	0.0	Do NOT use — distorts defect appearance

Dataset size rules of thumb

Images per class	Strategy
< 100	Heavy augmentation + copy_paste; consider anomaly detection instead
100–500	Standard fine-tune with copy_paste=0.4
500–2000	Standard fine-tune, reduce augmentation slightly
> 2000	Can use larger model (yolov8s-seg)

---

Phase 4: Evaluation

# After training
model = YOLO("runs/defect/yolov8n_seg_v1/weights/best.pt")
metrics = model.val(data="data.yaml", iou=0.5, conf=0.25)
print(metrics.seg.map)      # mAP50-95 for segmentation
print(metrics.seg.map50)    # mAP50
print(metrics.box.map50)    # bbox mAP50

Target metrics for industrial inspection:

• mAP50 > 0.85 → acceptable
• mAP50 > 0.92 → production-ready
• mAP50 < 0.75 → revisit data quality / add samples

Per-class analysis matters more than overall mAP for industrial use. A rare but critical defect class at 0.5 mAP is a problem even if overall mAP looks fine.

---

Phase 5: Export for Deployment

# TensorRT (NVIDIA GPU / Jetson)
model.export(format="engine", half=True, device=0)

# ONNX (cross-platform)
model.export(format="onnx", opset=12, simplify=True)

# TFLite (mobile/embedded)
model.export(format="tflite", int8=True)

---

Reference Files

• references/grayscale-finetune.md — Complete script for first-layer weight averaging + grayscale training pipeline. Read for any 1-channel / TIF input.
• references/training-config.md — Full data.yaml template, hyperparameter file, augmentation presets, and multi-GPU setup. Read before writing training code.