Dual-Strategy Improvement of YOLOv11n for Multi-Scale Object Detection in Remote Sensing Images

1. Problem Statement

The paper addresses the significant challenges associated with object detection in high-resolution remote sensing imagery (satellite and aerial). Despite advancements in deep learning, existing models struggle with three primary issues specific to this domain:

• Small Object Detection: Targets often occupy very few pixels (e.g., <32×32) due to long imaging distances, making feature extraction difficult.
• Complex Backgrounds & Low Contrast: Diverse scenes (urban, farmland, terrain) create high background clutter and low target-background contrast, leading to feature confusion.
• Extreme Multi-Scale Variation: Objects in a single image can range from vehicles (few pixels) to large buildings (hundreds of pixels). Traditional Feature Pyramid Networks (FPN) often fail to simultaneously preserve fine-grained spatial details and high-level semantic information across such a wide scale span.
• Limitations of Baseline Models: While YOLOv11n is lightweight and efficient, it lacks sufficient global contextual modeling and cross-scale feature interaction capabilities required for these complex remote sensing scenarios.

2. Methodology

The authors propose two distinct, lightweight improvement strategies based on the YOLOv11n architecture. Both strategies aim to enhance feature representation and fusion while maintaining real-time inference capabilities.

Core Components Introduced:

1. Large Separable Kernel Attention (LSKA):
   • Function: Integrated into the Backbone.
   • Mechanism: Decomposes large 2D convolution kernels into separable 1D convolutions (horizontal/vertical) combined with depthwise and dilated convolutions.
   • Goal: Captures large receptive fields and global contextual information to distinguish small objects from complex backgrounds without significantly increasing computational cost.
2. Gold-YOLO Neck (Gather-and-Distribute):
   • Function: Replaces the standard neck in both strategies.
   • Mechanism: Utilizes a "Gather-and-Distribute" (GD) structure with Low-GD (preserving spatial details for small objects) and High-GD (enhancing semantic context for large objects) branches. It includes Feature Alignment Modules (FAM) and Information Fusion Modules (IFM) to align and fuse features across different scales.
   • Goal: Achieves robust multi-scale feature fusion and improves information flow between shallow and deep layers.
3. MultiSEAMHead:
   • Function: Replaces the standard detection Head.
   • Mechanism: Incorporates depthwise separable convolutions, cross-layer connections, and channel/spatial mixing units. It uses attention-based modulation to reweight features.
   • Goal: Enhances the joint modeling of spatial and channel dependencies, improving the detection of occluded and dense objects.

The Two Proposed Models:

• Model 1: YOLOv11n-LSKA-GoldYOLO
  • Architecture: Backbone (LSKA) + Neck (Gold-YOLO) + Standard Head.
  • Focus: Enhances global context perception in the backbone and multi-scale fusion in the neck. Ideal for scenarios requiring strong background suppression and small object detection.
• Model 2: YOLOv11n-GoldYOLO-MultiSEAMHead
  • Architecture: Backbone (Standard) + Neck (Gold-YOLO) + Head (MultiSEAMHead).
  • Focus: Enhances feature interaction in the neck and strengthens the detection head's ability to handle dense, occluded, and multi-scale objects through advanced attention mechanisms.

3. Key Contributions

• Dual-Strategy Framework: Proposes two complementary lightweight models tailored for different remote sensing challenges (global context vs. dense/occluded detection) rather than a single monolithic solution.
• LSKA Integration: Successfully adapts the Large Separable Kernel Attention mechanism to the YOLOv11n backbone, solving the trade-off between large receptive fields and computational efficiency.
• Gold-YOLO Adaptation: Effectively integrates the Gold-YOLO gather-and-distribute mechanism into the YOLOv11n neck to address the multi-scale variation inherent in aerial imagery.
• Advanced Head Design: Introduces the MultiSEAMHead to improve the final regression and classification stages, specifically targeting dense and occluded targets.
• Lightweight Preservation: Both models maintain the real-time inference speed and low computational overhead of the original YOLOv11n, making them suitable for edge deployment.

4. Experimental Results

• Dataset: Experiments were conducted on the DOTA-v1 dataset (2,806 high-resolution images, 15 object categories), a standard benchmark for aerial image detection.
• Metrics: Evaluated using Precision, Recall, mAP@0.5 (IoU=0.5), and mAP@0.5:0.95.
• Performance Gains:
  • YOLOv11n-LSKA-GoldYOLO: Achieved a 1.3% increase in mAP@0.5 (from 42.0% to 43.3%) and a 0.3% increase in mAP@0.5:0.95 compared to the baseline.
  • YOLOv11n-GoldYOLO-MultiSEAMHead: Achieved a 1.8% increase in mAP@0.5 (from 42.0% to 43.8%) and a 0.6% increase in mAP@0.5:0.95.
• Qualitative Analysis: Visual results showed significant improvements in detecting small, sparse targets and handling dense clusters/occlusions compared to the baseline YOLOv11n and other YOLO versions (v8, v9, v10).

5. Significance

• Practical Application: The proposed models offer a flexible solution for remote sensing applications such as urban governance, smart agriculture, and national defense. Users can select Model 1 for tasks requiring global context (e.g., sparse vehicle detection) or Model 2 for complex, dense environments (e.g., port or city center monitoring).
• Efficiency vs. Accuracy: The work demonstrates that significant accuracy gains in remote sensing do not necessarily require heavy, computationally expensive models. By strategically modifying specific modules (Attention, Neck, Head), the authors achieved robust performance while retaining the lightweight nature of YOLOv11n.
• Future Direction: The study highlights the importance of adaptive multi-level fusion and attention mechanisms in overcoming the specific limitations of remote sensing imagery, providing a blueprint for future lightweight detector designs in this domain.