Pcb 348

PCB Anomaly Detection (Spatial PaDiM with Photometric Inputs)

Authors: Rikke Aa. Boysen, Simone I. Lebech, Andreas S. From
Supervisor: Jakob Wilm
Course: Semester Project (Advanced Robotics) Programme: M.Sc. in Advanced Robot Technology
Institution: University of Southern Denmark (SDU)
Semester: 2nd Semester, M.Sc. in Advanced Robot Technology, Spring 2026

This repository implements Spatial PaDiM (Patch Distribution Modeling) for anomaly detection on printed circuit boards (PCBs) using photometric (multi-illumination) image stacks. The framework supports synthetic photometric datasets with 24 fixed illumination directions by default and provides tools for dataset generation, model training, experimentation, and evaluation.

Key Features

• Inputs consist of grayscale photometric image stacks (24 illumination directions by default, resulting in 24 input channels). Binary illumination masks are not included in the input representation.
• Anomaly detection is based on a Spatial PaDiM approach. A modified ResNet-18 feature extractor processes multi-channel inputs, feature embeddings are collected at each spatial location, and a Gaussian distribution with Ledoit-Wolf covariance regularization is fitted per spatial position.
• Evaluation is performed using spatial Mahalanobis distances with top-k aggregation to produce PCB-level anomaly scores and corresponding ROC/AUC metrics.

Repository Status

This README reflects the current implementation. Model training is performed through src/train.py or src/run_experiment.py, dataset generation through src/experiment.py, and evaluation through src/evaluate.py.

Installation

1. Clone the repository.
2. Install the required dependencies:

pip install -r requirements.txt

Note: The project was developed using recent Python and PyTorch versions. Ensure that the installed PyTorch build is compatible with your CUDA version or CPU environment.

Dataset Structure

The dataset should be organized as follows:

dataset/
├── normal/
│   ├── pcb001/
│   │   ├── image01.png
│   │   ├── ...
│   │   └── image24.png
│   └── ...
└── anomaly/
    ├── pcb101/
    │   ├── image01.png
    │   ├── ...
    │   └── image24.png
    └── ...

Each PCB sample is stored in a separate folder containing 24 grayscale images, one for each illumination direction. By default, image files are expected to follow the naming convention image01.png through image24.png (see src/dataset.py and src/experiment.py).

Synthetic Dataset Generation

The synthetic data generator creates photometric image stacks using the illumination directions specified in src/blenderproc/lights_export.csv.

python src/experiment.py --num_normal 50 --num_anomaly 50

This command generates the dataset/normal/ and dataset/anomaly/ directories along with the corresponding PCB samples.

Model Training

Single Configuration

To train a model using the default configuration:

python src/train.py

The script invokes src.anomaly_model.train() using a predefined configuration suitable for quick experimentation.

Experiment Sweeps

For systematic parameter studies, use:

python src/run_experiment.py

This script executes multiple PaDiM configurations by varying parameters such as feature_dim, topk_ratio, and use_patches. Results are written to the experiment_logs/ directory.

Model Evaluation

Trained models can be evaluated using:

python src/evaluate.py --model_path "src/Training output/padim_model.pth" --save_heatmaps

Evaluation outputs, including CSV files, performance plots, and anomaly heatmaps, are saved to the training output directory specified by TRAINING_OUTPUT_DIR (src/Training output by default).

Repository Overview

src/dataset.py

Implements the PCBAnomalyDataset class responsible for loading photometric image stacks. Optional preprocessing steps include CLAHE enhancement and patch extraction.

src/experiment.py

Contains the synthetic PCB generation pipeline, including surface generation and photometric rendering through render_photometric_images().

src/anomaly_model.py

Implements the modified ResNet-18 feature extractor, Spatial PaDiM training (manual_padim_train), and inference (manual_padim_predict). Trained models are stored as padim_model.pth within the experiment output directory.

src/run_experiment.py

Provides automated execution of multiple experimental configurations and records results in the experiment_logs/ directory.

src/evaluate.py

Loads trained checkpoints and computes evaluation metrics such as ROC and Precision–Recall curves while generating qualitative visualizations.

Limitations

• Inputs consist of multi-channel grayscale photometric image stacks (24 channels by default). Binary illumination masks are not incorporated into the input representation.
• The framework implements Spatial PaDiM using Gaussian feature distributions and Mahalanobis distance-based anomaly scoring. Reconstruction-based methods such as autoencoders are not included.
• Some evaluation routines assume that backbone parameters are available in the saved checkpoint. When using custom checkpoints, verify compatibility between src/anomaly_model.py and src/evaluate.py.

Dependencies

The required packages are listed in requirements.txt.

Core dependencies include:

• PyTorch (torch)
• TorchVision (torchvision)
• NumPy (numpy)
• Pandas (pandas)
• Matplotlib (matplotlib)
• Scikit-learn (scikit-learn)