Object detection in aerial images poses significant challenges for computer vision. This is mainly due to the very small size of the targets, which occupy very few pixels, resulting in limited available feature information and making accurate recognition and classification complex. Moreover, the dense distribution of objects often leads to overlapping, further complicating precise classification and localization. Meanwhile, traditional detection networks struggle to strike a balance between improving recognition accuracy and managing model complexity. To address these issues, this paper proposes MFF-YOLO, a small object detection algorithm in aerial images based on multi-level feature fusion. First, we design a Multi-Level Feature Fusion module (MFF), which fuses features from multiple levels to mitigate feature loss caused by downsampling. Then, we combine Soft_NMS and NWD to propose the NWD_Soft_NMS algorithm, which replaces the non-maximum suppression algorithm (NMS) in the YOLOv8 network, thereby improving the detection performance for occluded objects. Experimental results show that compared to the baseline, the proposed algorithm achieves a 9.0% improvement in mAP and outperforms other algorithms on the VisDrone dataset.