Data Augmentation Using Image Generation for Change Detection

Learning-based change detection (CD) in water scenarios is a key functionality for unmanned aerial vehicle (UAV). However, computer vision algorithms require large number of labeled datasets. Inspired by parallel intelligence, we propose a systematic framework for data generation. In this work, the framework consists of simulated scene and image generation network. In simulated scene, simulated images with pixel-level annotations are automatically generated. Then, image generation network uses paired images (real and simulated) to generate synthetic images. We use simulated and synthetic images in combination with publicly available real-world images to conduct experiments. The experimental results indicate that: 1) simulated images can be used in change detection research; 2) synthetic images effectively improve the performance of supervised change detection model.