Objective Trackside inspection of train wheel treads is essential for railway safety. In real scenarios, wheel-tread defects often exhibit large-scale variations and weak visual saliency, while the acquired images are frequently degraded by complex background textures, uneven illumination, and noise. These factors increase missed detections and false alarms, particularly for small and subtle defects. To address these issues, this paper proposes an improved detector, termed MSCF-YOLO, for robust wheel-tread defect detection via multi-frequency feature perception and adaptive noise suppression. Specifically, the proposed method enhances backbone representation learning by integrating star operations into the C3k2 module, strengthens the joint aggregation of low- and high-frequency cues using a multi-frequency feature-space pyramid module, and introduces an SOSim module to amplify responses to small defects while adaptively suppressing background noise and clutter-induced interference.

Methods To achieve a favorable trade-off between detection accuracy and real-time performance for typical wheel-tread defects (pitting, spalling, and wear), MSCF-YOLO (Fig.1) is developed based on the lightweight YOLOv11n baseline and incorporates three targeted enhancements. 1) C3k2-SCAA (Fig.4) for multi-scale representation. The C3k2 block is restructured by embedding a StarNet-inspired feature mapping together with Context Anchor Attention (CAA) (Fig.2), forming Star-CAA-Block (Fig.3) to replace the original internal bottleneck. This design strengthens cross-layer semantic aggregation and mitigates feature attenuation during multi-scale fusion. 2) MBFSPPF (Fig.5) for multi-frequency feature fusion. A Multi-Band Feature Spatial Pyramid Pooling-Fast (MBFSPPF) module is designed by replacing fixed max-pooling with multi-dilation convolutions, which reduces spatial detail loss, enhances scale-aware context modeling, and enables complementary integration of low-frequency semantics and high-frequency details, thereby improving localization stability. 3) SOSim (Fig.9) for noise-robust small-defect enhancement. A Small-Object SimAM (SOSim) module is introduced to perform self-adaptive noise suppression, enhancing defect-related high-frequency responses while attenuating background clutter and noise interference, thus improving the effective signal-to-noise ratio for subtle defects.

Results and Discussions Experimental comparisons were conducted on the wheel-tread defect dataset under identical settings to demonstrate the effectiveness of the proposed method. The improved MSCF-YOLO was evaluated against representative detectors, including YOLOv5n, YOLOv7-tiny, YOLOv8n, YOLOv10n, YOLOv11n, RT-DETR, Mamba, LSKNet, EfficientViT, RepViT, ConvtextV2 and Swin-Tiny. As summarized in Tab.5, MSCF-YOLO achieves the best overall performance, attaining an mAP@50 of 87.5%, which ranks first among all compared methods. Notably, while delivering high detection accuracy, MSCF-YOLO maintains a compact model size and moderate computational cost, with 3.01 M parameters and 7.9 G FLOPs, which is comparable to YOLOv8n in magnitude. Moreover, MSCF-YOLO surpasses YOLOv8n in per-class detection accuracy and improves precision and recall by 3% and 2%, respectively. Although YOLOv5n provides the most lightweight configuration (1.76 M parameters and 4.9 G FLOPs), MSCF-YOLO leverages the collaborative module design to boost mAP@50 by 4.6%, with per-class gains of 0.3%, 13.2%, and 7.1%, respectively. To visually demonstrate the improvement over the baseline, qualitative comparisons before and after the proposed enhancements are presented in Fig.16. The baseline YOLOv11n tends to miss small defects, resulting in frequent false negatives, whereas MSCF-YOLO can reliably detect subtle defects that are missed by the baseline and yields more consistent localization and recognition, particularly for spalling and wear, thereby reducing missed detections and improving overall robustness.

Conclusions This paper presents an improved YOLOv11n-based detection framework for railway wheel-tread defect inspection. The proposed method integrates three efficient enhancement modules to improve detection accuracy. To strengthen cross-scale feature acquisition across network layers, StarNet-inspired star computation and Context Anchor Attention are incorporated into the C3k2 block, yielding an enhanced C3k2-SCAA module. In addition, the proposed MBFSPPF replaces the fixed max-pooling operations in the conventional SPPF module with multi-dilation convolutions, and stacks CBAM and CAFM attention mechanisms to reinforce channel-wise fusion of multi-dimensional feature maps while reducing spatial information loss during feature aggregation. Furthermore, an SOSim module is inserted before the detection head to enhance small-defect sensitivity and suppress the adverse impact of background noise. Compared with the baseline YOLOv11n, MSCF-YOLO achieves mAP@50=87.5% and mAP@50–95=50.1%, corresponding to improvements of 3.7% and 4.0%, respectively. It further attains per-class AP@50 scores of 95.7%, 78.0%, and 89.0% for the three defect categories, with corresponding gains of 0.8%, 7.5%, and 3.0%, respectively. Although MSCF-YOLO increases the parameter count by 16% relative to YOLOv11n, it still maintains a high inference throughput of 131.8 FPS, satisfying real-time requirements. These results demonstrate the practicality and effectiveness of MSCF-YOLO for railway wheel-tread defect detection. Future work will explore knowledge distillation and model pruning to remove redundant weights and further improve detection efficiency while preserving accuracy, thereby reducing missed and false detections in real-world applications.