Enhanced Anomaly Detection in Aerial Imagery using Text to Image Generation

This study investigates the use of generative AI models, like unCLIP for synthetic aerial image generation to improve anomaly detection in the field of aerial imagery. The research demonstrates how unCLIP can generate realistic aerial images from textual descriptions, addressing challenges related to the limited availability of real-world data. These artificially generated images with anomalies are then used to train and optimize the YOLOv8 model for anomaly detection, with a focus on detecting environmental and infrastructural anomalies. The study explores the entire process—from prompt design and image generation to data preprocessing, anomaly labeling, and model training—while evaluating the effectiveness of synthetic data qualitatively and quantitatively in enhancing model performance. While the results indicate that the generated images significantly improve anomaly detection capabilities, challenges in achieving high image fidelity, accuracy and realism still exist. All the findings of this research work underscore the potential of synthetic data generation in training robust machine learning models and provide a foundation for future research to improve the scalability, accuracy, and generalizability of anomaly detection systems.