MFSS-YOLO for small object detection in remote sensing images

In recent years, many studies have successfully leveraged convolutional neural networks to improve the performance of In recent years, many studies have successfully leveraged convolutional neural networks for enhancing object detection performance in remote sensing images. However, the characteristics of optical remote sensing imaging pose several challenges, such as the small size of objects, which complicates feature extraction. Additionally, optical remote sensing images are often affected by complex background information and adverse weather conditions, which heighten the challenge of distinguishing objects from the background. To address these issues, this paper proposes an efficient small object detection model for optical remote sensing, called Multi-Scale Feature Fusion YOLO (MFSS-YOLO), to enhance detection accuracy and reduce the model's parameter count. MFSS-YOLO consists of two modules: the Replicated Hybrid Global Stem (RepHGStem) module and the Multi-Scale Feature Fusion (MFSS) module. The RepHGStem enhances the model's capabilities for contextual information capture and feature extraction by expanding the receptive field of the stem. The MFSS module further enhances the model's multi-scale feature extraction capability. Extensive experiments were conducted on two public small object datasets (AI-TODv2 and TinyPerson), and the results demonstrate that MFSS-YOLO achieves satisfactory performance in terms of both efficiency and accuracy, showing advantages in detection accuracy and model efficiency.