NiSA: An improvement method for anomaly detection using synthetic data

Introduction

Based on NSA (Natural Synthetic Anomaly), a self-supervised model for anomaly detection, we propose a new model after modifying its anomaly data synthesis method, anomaly score calculation and deep learning network architecture, We proposed a new end-to-end model - NiSA (Natural improved Synthetic Anomaly). Experiments on MVTec AD (mainstream dataset for anomaly detection, containing images of 15 different industrial objects) show that NiSA can produce more natural, wide-ranging synthetic anomalies and higher overall detection AUROC.

Method

1. Generates a Perlin noise mask of the same size as the rectangular patch and superimposes the matrix on rectangular shape patches, the restrictions are as follows:

2. Generates a Perlin noise mask of the same size as the rectangular patch and superimposes the rectangular shaped patches, which are restricted as follows:

3. Choose an anomaly generation method (blur, scar, hole, clone) and generate the anomaly and paste in the target location.

4. Repeat steps 1~3 until the abnormal area generated is sufficient

5. Create a corresponding pixel-wise label mask from the pasted anomaly patch

6. Generated synthetic data

Previous Next

Workflow

The network model of NiSA is an encoder-decoder architecture based on ResNet-50, ResNet-50 removes the last classification layer as an encoder, 4 1/2 down sampling 1x1 convolutional layers as a The network model of NiSA is an encoder-decoder architecture based on ResNet-50, ResNet-50 removes the last classification layer as an encoder, four 1/2 down sampling 1x1 convolutional layers as a bottleneck plus a simple decoder, And add the attention module of SSPCAB to achieve the purpose of enhancing featuresbottleneck plus a simple decoder, And add the attention module of SSPCAB to achieve the purpose of enhancing features.

Result

Testing Result

Detection AUROC

Segmentation AUROC

Ablation Study

Previous Next