MFP-YOLO: A Multi-scale Feature Fusion and Perception-Enhanced YOLOv11 for Remote Sensing Small Object Detection

Abstract

Small-object detection in remote-sensing imagery remains challenging because targets often occupy only a few pixels and are further affected by substantial scale variations, dense spatial distributions, and complex backgrounds. To address these issues, this study develops MFP-YOLO, a lightweight detector based on YOLOv11, by introducing targeted improvements in feature extraction, feature fusion, and bounding-box regression. First, the Multi-Features Extraction Block (MFEB) is incorporated into C3k2 to construct C3K2-MFEB, which introduces progressive multi-scale receptive-field modeling and adaptive scale selection while preserving the cross-stage partial feature reuse mechanism of YOLOv11. Second, the neck is reorganized as a Modulated Small-object Enhancement Pyramid (MSEP), which combines small-object feature enhancement with modulation-based feature fusion to strengthen the transmission of spatial details and adaptively regulate the contributions of multi-level features. Finally, Focaler-PIoU (FPIoU) integrates IoU-range focusing with a boundary-distance penalty to improve the localization accuracy of densely distributed small objects. Experimental results on the RSOD dataset show that MFP-YOLO achieves an mAP@0.5 of 96.75\% with only 3.02 M parameters and 9.3 GFLOPs, representing a 4.32\% improvement over the YOLOv11 baseline while maintaining an inference speed of 152.3 FPS. Further evaluation on the SIMD dataset also confirms the effectiveness of the proposed framework. Overall, the results demonstrate that the integrated modifications provide a favorable trade-off between detection accuracy and computational efficiency for remote-sensing small-object detection.We have uploaded the core code to https://github.com/yuanshicha/ArticleCode.git.