You Only Look Once with Clustering Segmentation Module

Detecting objects in aerial images presents considerable difficulties because of the following reasons: 1) Aerial images are often very large, consisting of millions or even hundreds of millions of pixels, and these images pose significant challenges due to limited computational resources. 2) The small size of objects results in a lack of sufficient information for accurate detection. 3) The uneven distribution of objects causes inefficient use of computational resources. To tackle these problems, we propose YOLOCSM (You Only Look Once with Clustering Segmentation Module), an efficient and effective framework. To address the difficulties arising from large-scale images and uneven object distribution, we propose a Clustering Segmentation Module (CSM), which dynamically identifies cluster regions and zooms in on them to enhance detection accuracy. Additionally, we propose a new training strategy to optimize small targets and intensive targets in complex scenarios. We conducted extensive experiments on the UAVDT aerial image dataset and compared our method with the state-of-the-art model YOLOv11, demonstrating the effectiveness and superiority of the proposed approach.