WASP PhD - ML for anomaly detection
Title: WASP PhD - ML for anomaly detection
SNIC Project: Berzelius-2021-65
Project Type: LiU Berzelius
Principal Investigator: Adha Hrusto <adha.hrusto@cs.lth.se>
Affiliation: Lunds universitet
Duration: 2021-10-30 – 2022-03-01
Classification: 10201
Keywords:

Abstract

In this project, we aim to develop, evaluate and analyze pixel-wise anomaly detection (AD) in images of primarily natural environments. Anomalies in this sense are pixels or patches of pixels that have different spectral or textural properties compared to the background image and could stem from human-made objects. The aim scenario of this AD lies in search-and-rescue missions and is thus connected to the WASP Research Arena for Public Safety (WARA PS). We address AD from two directions with (i) a parametric, probabilistic method and (ii) an image reconstruction method. In (i) we will perform AD in color space using the Reed-Xiaoli Detector (RX), whereas in (ii) we will train a generative adversarial neural network (GAN) to generate background images (reconstructions) without anomalies to detect differences between original and reconstructed images as anomalies. These differences could be measured either in the image or the feature space. For training the GAN, we will investigate publicly available datasets as well as create our own dataset using high-resolution aerial orthophotos. Both approaches will be quantitively evaluated on the publicly available or self-made datasets and qualitatively evaluated on recordings from unmanned aerial vehicles (UAV) provided through WARA PS.