[PyTorch] Reconstruction by inpainting for visual anomaly detection (RIAD)

PyTorch implementation of "Reconstruction by inpainting for visual anomaly detection (RIAD)"

Concept

Concept ot the RIAD [1].

Model

Training Strategy

Overal Procedure

Preprocessing

Inference and Postprocessing

Experiments

Preparing for Disjoint Masking

Input (Given)

Set of disjoint masks

Disjoint masks with $k={2, 4}$.
First and second row shows the mask of $k=2$ and $k=4$ respectively.
Each column shows the $i$-$th$ mask for each cell size $k$.

Set of mask applied input

Anomaly Detection using MNIST dataset

Setting

• Normal (Good): 1
• Abnormal (Not-good): 0, 2, 3, 4, 5, 6, 7, 8, 9 (other than 1)

Results

Loss convergence

Reconstruction

Anomaly detection performance (w/ validation set)

Result of training: result.json (w/ test set)

root

name_best:"model_2_best_auroc.pth"
auroc:0.9974551381667722
loss:0.0017763811201996548
select_norm:1
masking_mode:"disjoint_mask"
disjoint_n:3
nn:2000
dim_h:28
dim_w:28
dim_c:1
ksize:3
mode_optim:"adam"
learning_rate:0.001
mode_lr:0
path_ckpt:"Checkpoint"
ngpu:1
device:"cuda"
filters:"[1, 64, 128, 256, 512]"

Requirements

• PyTorch 1.11.0

Reference

[1] Vitjan Zavrtanik et al. "Reconstruction by inpainting for visual anomaly detection." Pattern Recognition, vol. 112, 2021.
[2] Taiki Inoue. MSGMS (python module).