AERIAL IMAGE SEMANTIC SEGMENTATION USING U-NET AND EFFICIENT-NET

Abstract

To address the challenges of few-shot aerial image semantic segmentation, where unseen-category objects in query aerial images need to be parsed with only a few annotated support images, we propose a novel approach by integrating a UNet architecture with Efficient Net. The rotation sensitivity of aerial images causes substantial feature distortions, leading to low confidence scores and misclassification of same-category objects with different orientations. To overcome these limitations, we propose an enhanced solution by combining U-Net for robust semantic segmentation with Efficient Net for efficient feature extraction and scale adaptability. This architecture, which we refer to as Efficient U-Net, introduces rotation-invariant feature extraction to handle the varying orientations of aerial objects. By leveraging Efficient Net’s scalable Convolutional layers for feature extraction, we ensure that the network can capture orientation-varying yet categoryconsistent information from support images. This approach enhances the segmentation accuracy by aligning same-category objects, irrespective of their orientation, thereby minimizing the oscillation of confidence scores and improving the detection of rotated semantic objects. This Efficient U-Net model provides a scalable, rotation-invariant solution to the few-shot segmentation of aerial images.