MS-YOLOv11: A Wavelet-Enhanced Multi-Scale Network for Small Object Detection in Remote Sensing Images

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Small objects (<32x32 pixels) in remote sensing imagery present detection challenges due to weak signals, limited features, and information loss during down-sampling.
• Existing object detectors struggle with high miss rates and inadequate feature extraction for diminutive objects.

Purpose of the Study:

• To develop an enhanced object detection model, MS-YOLOv11, specifically designed to address the challenges of small object detection in remote sensing.
• To improve the accuracy, efficiency, and feature representation for detecting small objects.

Main Methods:

• Proposed MS-YOLOv11, integrating frequency-domain detail preservation using 2D Haar wavelet decomposition.
• Employed lightweight receptive-field expansion via small-kernel depthwise convolutions on sub-bands.
• Utilized Mix Structure Block (MSB) with MSPLCK and EPA modules for adaptive cross-scale fusion and feature re-weighting.

Main Results:

• MS-YOLOv11 demonstrated superior performance on DOTA and DIOR datasets compared to the baseline.
• Achieved improvements in mean Average Precision (mAP@50, mAP@95), parameter efficiency, and inference speed.
• Validated the model's effectiveness in targeted small-object detection.

Conclusions:

• MS-YOLOv11 effectively counteracts challenges in small object detection by preserving high-frequency details and adaptively fusing multi-scale features.
• The proposed methods enhance discriminative feature extraction and suppress background interference for improved detection accuracy.