Attention Guided Anomaly Localization in Images

Paper link: https://arxiv.org/pdf/1911.08616.pdf

Proposed system: Adversarial Variational autoencoder with Guided Attention (CAVGA), which localizes the anomaly with a convolutional latent variable to preserve the spatial information. The two settings performed here are with and without the need of anomalous training images.

Challenge: small pixel coverage of the anomaly and limited sample size of the anomalous images.

Using motivation from GAIN , attention based supervision can alleviate the need for large training data and localizes anomalies without using any anomly data.

Main Idea: given only normal images, the attention map is asked to focus on all the normal regions where as decrease the attention in the anomly part. During testing, once the image is detected as anomalous, the attention map localizes the anomaly in the picture.

In unsupervised setting : attention expansion los is used.

In weekly supervised setting, : complementary guided attention loss is used.