Scientific Reports - Article in Press

YOLOv11-SRA: An Adaptive & Robust Detection Framework

To overcome the inherent challenges of underground safety helmet detection, this paper proposes the YOLOv11-SRA model, an innovative extension of the YOLOv11 architecture. This model integrates a three-stage optimization strategy designed for dynamic and robust performance:

• SAConv (Switchable Atrous Convolution) Module: Replaces traditional convolution layers to dynamically adjust dilation rates, effectively capturing multiscale contextual information and enhancing the robustness of small target detection, particularly in varying lighting conditions.
• RCM (Rectangular Self-Calibration Module): Optimizes the neck network by employing rectangular self-calibrated attention. This mechanism refines foreground regions and significantly improves the model's boundary localization capabilities, crucial for detecting partially occluded helmets.
• ASFF (Adaptive Spatial Feature Fusion) Module: Enhances the detection head by fusing multiscale features through adaptive spatial weighting. This strategy alleviates feature conflicts and improves the accuracy of small safety helmet detection, especially in low-resolution and dense scenes.

The synergistic integration of these modules enables YOLOv11-SRA to achieve closed-loop optimization from feature extraction to fusion to detection, effectively addressing scale changes, occlusion interference, and low contrast levels in harsh underground environments.

Core Technical Innovations

The YOLOv11-SRA model is built upon three key innovations that provide dynamic adaptability and precision:

• SAConv: Dynamic Receptive Field Adaptation
  The Switchable Atrous Convolution (SAConv) module addresses the diverse receptive field requirements of multiscale objects. It constructs an adaptive feature extraction structure by dynamically selecting optimal dilation rates (e.g., 1 for local details, 5 for long-range context) through a learnable switching mechanism. This enables efficient capture of multiscale features in underground scenes.
• RCM: Geometric Foreground Refinement
  The Rectangular Self-Calibration Module (RCM) is integrated into the neck network to enhance foreground localization. Unlike conventional square kernels, RCM uses axial pooling to extract global context and large-kernel striped convolutions to calibrate attention regions. This aligns attention more effectively with the elongated geometric patterns of partially visible helmets, improving boundary accuracy even with occlusions.
• ASFF: Adaptive Multiscale Feature Fusion
  The Adaptive Spatial Feature Fusion (ASFF) module optimizes the detection head by dynamically learning spatial adaptive weights. This mechanism selectively fuses features from different scale layers, mitigating conflicts and improving the accuracy of small target detection. It ensures that only pertinent information is combined, effectively addressing blurred edges and enhancing localization in dense, low-resolution images.

Performance & Validation

The YOLOv11-SRA model was rigorously validated on the CUMT-HelmeT dataset, demonstrating significant improvements over mainstream object detection models in both accuracy and efficiency.

Quantitative Performance Comparison (mAP50)

Ablation Study: Module Contributions (mAP50)

The ablation study confirms the synergistic effect of SAConv, RCM, and ASFF. While RCM alone showed a performance decline due to its need for rich contextual information, its integration with SAConv's dynamic scaling and ASFF's adaptive fusion resolves these issues, leading to the overall best performance.

Qualitative Improvements in Challenging Scenarios