LiMR

Lightweight Masked Reconstruction for Real-Time Sensor-Driven Anomaly Detection in Industrial IoT

🌐 中文

📄 Paper  |  ⚡ C++ Implementation

---

• LiMR
• 📊 Performance
  • AeBAD-S (Aero-Engine Blade)
  • MVTec AD
  • Jetson AGX Xavier Edge Deployment
• 🚀 Quick Start
  • ⌨️ Training
  • ⌨️ Testing
• 📦 Supported Datasets
• ⚙️ Parameter Reference
  • Data Parameters
  • Model Parameters
  • Training Parameters
• 📜 Batch Scripts
• 📂 Output Structure
• ⚡ TensorRT High-Speed Deployment
• ❓ FAQ
• 📚 Reference
  • Acknowledgements

---

This is the anomalib-integrated implementation of LiMR, a lightweight teacher-student architecture for visual anomaly detection. The model uses a frozen teacher encoder to extract multi-scale features, while a lightweight MobileViTv2-based student encoder-decoder reconstructs semantic features under masked reconstruction — achieving competitive accuracy at high throughput.

📄 Original repository: ShowayLiao/LiMR 📘 Anomalib upstream README: originalREADME.md

---

📊 Performance

AeBAD-S (Aero-Engine Blade)

LiMR achieves state-of-the-art performance on AeBAD-S while drastically reducing model complexity.

Method	Params (M)	FLOPs (G)	Latency (ms)	Throughput (img/s)	AUROC (%)	PRO (%)
MMR	170.98	18.60	33.83	29.56	84.7	89.1
PatchCore	68.95	11.46	79.62	12.56	71.0	87.8
RD	91.75	31.61	25.98	38.49	81.0	85.6
SimpleNet	11.68	1.83	63.81	15.67	58.4	68.3
LiMR-175	40.06	10.21	23.26	42.99	85.3	91.1
LiMR-175 (TensorRT)	40.06	10.21	8.26	121.07	85.3	91.1

Compared to MMR: 76.6% fewer params, 45.1% lower FLOPs, 45.5% higher throughput, while surpassing MMR in both AUROC and PRO.

MVTec AD

LiMR maintains competitive accuracy on the standard MVTec AD benchmark.

Method	Image AUROC (%)	Pixel AUROC (%)
PatchCore	99.1	98.1
RD	98.5	97.8
SimpleNet	99.6	98.1
MMR	98.4	97.2
LiMR	97.5	96.9

LiMR achieves comparable accuracy to SOTA methods while using a lightweight CNN-ViT hybrid architecture — ideal for resource-constrained deployment.

Jetson AGX Xavier Edge Deployment

Method	VRAM (GB)	Latency (ms)	Throughput (FPS)	AUROC (%)
LiMR	1.44	82.22	12.16	85.32
LiMR (TensorRT FP32)	1.20	32.01	31.18	85.32
LiMR (TensorRT FP16)	0.48	15.66	63.87	85.23
MMR	1.64	106.32	9.41	84.73

TensorRT FP16 achieves 63.87 FPS real-time inference with only 0.48 GB VRAM on edge devices.

---

🚀 Quick Start

⌨️ Training

# Sync environment (first time only)
uv sync --extra cu126

# MVTec (single category)
uv run python tools\limr\train.py --dataset mvtec --root ./datasets/MVTec --category bottle \
    --image-size 224 --backbone resnet50 --alpha 1.75 --epochs 200

# AeBAD-S
uv run python tools\limr\train.py --dataset aebad_s --root I:\exp\datasets\AeBAD\AeBAD_S \
    --category AeBAD_S --image-size 256 --backbone resnet34 --alpha 1.75 \
    --fpn-output-dim 64 128 256 512 --block-dropout 0.0 --block-ffn-dropout 0.0 \
    --block-attn-dropout 0.0 --frozen-stages 3 --epochs 200 --seed 54

# AeBAD-V
uv run python tools\limr\train.py --dataset aebad_v --root I:\exp\datasets\AeBAD\AeBAD_V \
    --category AeBAD_V --image-size 256 --backbone resnet34 --alpha 1.75 \
    --fpn-output-dim 64 128 256 512 --block-dropout 0.0 --block-ffn-dropout 0.0 \
    --block-attn-dropout 0.0 --frozen-stages 3 --epochs 200 --seed 54

# RealIAD
uv run python tools\limr\train.py --dataset realiad --root ./datasets/RealIAD --category <category> \
    --realiad-resolution 256 --realiad-json ./datasets/RealIAD/RealIAD.json

# VISA
uv run python tools\limr\train.py --dataset visa --root ./datasets/VISA --category candle \
    --image-size 224 --backbone resnet50

# Folder (custom dataset)
uv run python tools\limr\train.py --dataset folder --root ./datasets/my_data \
    --folder-normal-dir normal --folder-abnormal-dir abnormal --folder-mask-dir mask \
    --category my_dataset

⌨️ Testing

# Test anomalib checkpoint (.ckpt)
uv run python tools\limr\test.py --dataset mvtec --root ./datasets/MVTec --category bottle \
    --image-size 224 --backbone resnet50 --alpha 1.75 --checkpoint ./output_limr/xxx.ckpt

# Test original LiMR weights (.pth)
uv run python tools\limr\test.py --dataset aebad_s --root I:\exp\datasets\AeBAD\AeBAD_S \
    --category AeBAD_S --image-size 256 --backbone resnet34 --alpha 1.75 \
    --block-dropout 0.0 --block-ffn-dropout 0.0 --block-attn-dropout 0.0 \
    --frozen-stages 3 --seed 54 --original-checkpoint I:\exp\LiMR\best_student_model_175.pth

📘 Note: Testing must use the same backbone, alpha, frozen-stages, fpn-output-dim, and block-*dropout settings as training, otherwise weight loading will be incomplete.

---

📦 Supported Datasets

--dataset	Dataset	Notes
mvtec	MVTec AD	15 industrial categories
mvtecad2	MVTec AD 2	Extended categories
mvtec_loco	MVTec LOCO	Logical constraints
btech	BeanTech	3 categories
bmad	BMAD	Biomedical anomaly
mpdd	MPDD	Metal parts
vad	VAD	Vehicle anomaly
visa	VISA	12 categories
realiad	RealIAD	30 categories, multi-resolution
kolektor	KolektorSDD2	Surface defect
aebad_s	AeBAD-S	Static images + 4 domain shifts
aebad_v	AeBAD-V	Video frames
folder	Folder	Custom dataset

---

⚙️ Parameter Reference

Data Parameters

Parameter	Description	Default
--dataset	Dataset name	mvtec
--root	Dataset root directory	./datasets/MVTec
--category	Sub-category / object class	bottle
--image-size	Input image size (square)	256
--train-batch-size	Training batch size	16
--eval-batch-size	Evaluation batch size	16
--num-workers	DataLoader workers	6

Model Parameters

Parameter	Description	Default
--backbone	Teacher encoder	resnet50
--alpha	Student width multiplier (1.0=100%)	1.75
--mask-ratio	Training mask ratio	0.4
--test-mask-ratio	Test mask ratio	0.0
--fpn-output-dim	FPN output channels per layer	auto-detect
--block-dropout	MobileViTBlockv2 dropout	0.1
--block-ffn-dropout	FFN dropout	0.0
--block-attn-dropout	Attention dropout	0.0
--frozen-stages	Frozen encoder stages (1-3)	3

Training Parameters

Parameter	Description	Default
--epochs	Max training epochs	200
--lr	Learning rate	0.001
--weight-decay	Weight decay	0.05
--warmup-epochs	LR warmup epochs	15
--early-stop-patience	Early stopping patience	10
--seed	Random seed	54

Test & Output Parameters

Parameter	Description
--checkpoint	anomalib Lightning ckpt path (.ckpt)
--original-checkpoint	Original LiMR weights path (.pth)
--output-dir	Output directory
--project-name	W&B project name

---

📜 Batch Scripts

Double-click any .bat file to run. Ensure .venv is created beforehand.

Script	Purpose
scripts\limr\train_aebad_s.bat	Train AeBAD-S
scripts\limr\train_aebad_v.bat	Train AeBAD-V
scripts\limr\train_mvtec.bat	Train MVTec (single category)
scripts\limr\train_mvtec_all.bat	Train MVTec (all 15 categories)
scripts\limr\train_visa.bat	Train VISA
scripts\limr\train_realiad.bat	Train RealIAD
scripts\limr\train_folder.bat	Train custom dataset
scripts\limr\test_aebad_s.bat	Test AeBAD-S
scripts\limr\test_aebad_v.bat	Test AeBAD-V
scripts\limr\test_mvtec.bat	Test MVTec
scripts\limr\test_visa.bat	Test VISA
scripts\limr\test_realiad.bat	Test RealIAD
scripts\limr\test_folder.bat	Test custom dataset

---

📂 Output Structure

After training, the following is generated under --output-dir:

output_limr/aebad_s/
├── LiMR/AeBAD_S/AeBAD_S/v34/
│   └── weights/
│       ├── lightning/model.ckpt   # Best checkpoint
│       └── onnx/model.onnx        # ONNX export
├── inference_speed.json           # Inference benchmark
└── wandb/                         # W&B logs (if enabled)

---

⚡ TensorRT High-Speed Deployment

After training, the ONNX model is auto-exported to weights/onnx/model.onnx under --output-dir. Use the LiMR C++ TensorRT repository for high-speed inference.

---

❓ FAQ

Out of memory (OOM)?

• Reduce --train-batch-size (e.g., 8 or 6)
• Reduce --num-workers
• Switch --backbone to resnet18

Weight loading errors during testing?

Ensure test parameters match training. When using --original-checkpoint to load legacy weights, you must use --backbone resnet34 (default in the original repo).

How does AeBAD multi-shift testing work?

AeBAD testing automatically iterates over all domain_shift subdirectories (same, background, illumination, view) and reports cross-shift average metrics at the end. Each shift's individual results are also logged to W&B.

---

📚 Reference

If you find LiMR useful in your research or work, please cite the original paper:

@article{liao2026limr,
  author  = {Shaowei Liao and Wenyong Yu and Shaolin Liao},
  title   = {Lightweight Masked Reconstruction for Real-Time Sensor-Driven
             Anomaly Detection in Industrial IoT},
  journal = {IEEE Internet of Things Journal},
  year    = {2026},
  doi     = {10.1109/JIOT.2026.3712733}
}

Acknowledgements

We acknowledge the excellent open-source implementations that this work builds upon:

• ConvMAE
• MobileViTv2
• MobileViTv2-pytorch
• MMR