Anomaly detection with sinusoidal representation network

Abstract

Anomaly detection in images helps to identify abnormal or unusual patterns in images relative to normal data. This issue is crucial in various domains, such as detecting abnormal areas in medical imaging, surface inspection, or photo editing. Many of the proposed methods require additional markup on the images. One way to solve this problem is image reconstruction. This method does not require additional markup on training data and trains on normal images. In this study, we present an approach that solves the Image Reconstruction problem by exploiting the Sinusoidal Representation Network (SIREN). SIREN is capable of modeling complex signals with great detail in small regions. We experimentally combined different loss functions for our architecture to improve the visual perception of the image. This study will also describe the different approaches we have tried in image reconstruction and our method’s evolution.