Anomaly Detection from Image Classification

This study aims to perform anomaly detection on human poses without using any pose estimator. Human pose anomaly detection has multiple practical applications across various industries and realms, contributing to improved safety, efficiency, and overall human-machine interaction. Here, we can also apply this in detecting anomaly poses of workers in a factory. To manifest this, we can come up with a system in which an image classifier is trained on a dataset of factory workers with various poses then deployed for pose anomaly detection. Contrary to previous works, it can be noted that the system can perform pose anomaly detection without any need for a separate pose estimator. Here, the definition of anomaly can vary by different datasets, so we use Stable Diffusion to generate a consistent dataset of normal and abnormal poses. Then, in order to achieve state-of-the-art performance, we use the EfficientNetV2 model for detecting anomalies within the generated dataset. Results show that the EfficientNetV2 model has excellent results on the dataset that is generated by Stable Diffusion.