Introduction

Goal

1. a pre-trained teacher network, a denoising student encoder-decoder, and a segmentation network in one framework.
   1. detection + location
2. Previous works focus on improving the similarity of S-T representations on normal inputs, whereas our work attempts to differentiate their representations on anomalous input.

Method

1. step (a), the student network with synthetic input is trained to generate similar feature representations as the teacher network from the clean image.
2. the teacher network is an ImageNet pre-trained ResNet18 with the final block removed (i.e., conv5 x).

step (b), the element-wise product of the student and teacher networks’ normalized outputs are concatenated and utilized to train the segmentation network.

Segmentation Network

1. [26,31], the cosine distances from multi-level features are summed up directly to represent the anomaly score of each pixel. However, the results can be suboptimal if discriminations of all level features are not equally accurate.
2. To address this issue, we add a segmentation network to guide the feature fusion with additional supervision signals.

Synthetic Anomaly Generation

Random two-dimensional Perlin noise

**DRÆM – A discriminatively trained reconstruction embedding for surface anomaly detection, iccv21**

Experiments

far better than [31] and PatchCore.

Failure cases of our method.