EMS-Net: Efficient Multiscale Perceptual Enhancement Tiny Object Detector for Remote Sensing Images

Abstract

Detecting tiny objects in remote sensing images has always been a challenging and intensive research area. This problem has not been well-solved due to the fact that object detection (OD) in remote sensing images is characterized by large-scale variations and complex backgrounds. On this basis, we propose the efficient multi-scale semantic-aware network (EMS-Net) constructed based on YOLOv8s for tiny OD network in remote sensing images. First, a new module multibranch context aggregation (MCA) is proposed to improve deep feature extraction and deep feature fusion of the model. In addition, we use our self-designed multiscale feature communication module (MFCM) aimed at reducing the loss of semantic information of object and mitigating the obstruction of foreground object by complex background. Finally, Wise IoU-Normalized Wasserstein distance (WIoU-NWD) is used as the bounding box regression loss to adapt the model to different object scale while improving the ability to localize tiny object. Comprehensive experiments on three popular datasets demonstrate that our method outperforms existing detectors, particularly in detecting tiny objects. Specifically, our approach achieves the mean average precision (mAP) of 77.2% on the DIOR dataset, 96.7% on the Remote Sensing Object Detection (RSOD) dataset, and 75.1% on the DOTA-v1.5 dataset.