Aerial Image Object Detection Based on Superpixel-Related Patch

Jiehua Lin, Yan Zhao, Shigang Wang, Meimei Chen, Hongbo Lin, Zhihong Qian

Abstract: Aerial image object detection and recognition has attracted increasing attention in recent years. Many excellent detectors have been proposed. However, due to the high-resolution of aerial images, these detectors are difficult to directly apply to aerial images. In order to solve the problem of hard processing caused by high resolution, it is generally to resize the high-resolution images into low-resolution images or cut the high-resolution images into small image patches. Cutting high-resolution aerial images into small image patches without overlap may cut an object into multiple parts which may lose the integrity of the object and causes one object to be detected as multiple objects. We design a new baseline to cut high-resolution aerial images into small image patches by using superpixel. Firstly, we use pixel-related GMM (Gaussian mixture model) to segment the high-resolution aerial images into superpixel images. Then we utilize superpixel label to cut high-resolution aerial images into low-resolution image patches with integrity of the object. Finally, we use YOLOv5 with CSL (Circular Smooth Label) to detect oriented objects. Our method effectively preserves the integrity of the object and improves the AP (Average Precision) of the object detection. This baseline can be applied not only to object detection, but also to aerial image segmentation, classification and so on. Experiments on the UCAS-AOD dataset show the effectiveness of the proposed method.