Leveraging GANs, autoencoders, and their combined deep learning architecture for anomaly detection in multi-domain image data

Abstract

The study aims to identify anomalies in multiple domains through deep learning and it evaluates the performance of 9 state-of-the-art models for anomaly detection in industrial quality control, plant disease diagnosis, and medical imaging. It compares them and proposes a modification to the Bi-GAN model because this modification is expected to improve the model’s performance. The 9 models are: FCAE, CAE, VAE, CVAE, Beta-VAE, VQ-VAE, AnoGAN, GANomaly, and Bi-GAN, and some are based on autoencoders (mirror) and others are based on GANs (adversarial competition between two agents). The authors modified the Bi-GAN model and they employed 3 datasets: brain MRI scans to detect abnormalities, industrial components to detect defects, and plant leaves to detect diseases, thus allowing them to test the models in different domains. The modified Bi-GAN model was the most effective model for identifying anomalies through the comparison of images, and this makes the model a good fit for anomaly detection in a variety of domains. All nine models showed promise, but each model also had its limitations, because by modifying the Bi-GAN model, the researchers were able to create a more robust and reliable approach.The study contributes and an evaluation and comparison of nine anomaly detection types. An introduction of a novel anomaly detection type that performs well in general, and a comparison of all anomaly detection types. This highlights the strengths and weaknesses of each type, and this provides good insights. The study contributes to our understanding of anomaly detection, and it also provides a basis for the design of novel anomaly detection types.