Enhancing small object detection in remote sensing imagery with advanced generative adversarial networks

This study underscores the significance of accurately identifying minute objects within satellite imagery. We propose an innovative GAN framework tailored for enhancing the detection of minute objects in such imagery. Our approach primarily emphasizes differentiating these objects from their surroundings by leveraging the core mechanisms of GANs. Moreover, we address the varied spectral characteristics of the background by integrating a similarity constraint into our GANs architecture. This strategy effectively distinguishes between the target object regions and non-object regions, significantly reducing the occurrence of false positives. Additionally, we employ a smoothness regularization technique to preserve the structural integrity of the detected objects. Through extensive experimentation on publicly available remote sensing datasets—spanning mineral, fabric, and vehicle detection—we compare our method quantitatively against conventional small object detection techniques and state-of-the-art GAN-based models. The comparative results consistently showcase the superior performance and robustness of our GAN-based approach in detecting minute objects within remote sensing imagery.