As the complexity of railway transportation systems increases, higher demands have been placed on the real-time performance and accuracy of track defect detection, making traditional methods inadequate for intelligent inspection tasks. To address this, this paper proposes a lightweight and efficient improved detection model, EADI-YOLO, based on the YOLOv10n framework, aiming to achieve the optimal balance between detection accuracy, inference speed, and computational cost. The model integrates the C2f-EMAN module into the backbone, combining EMA and NAM attention mechanisms to enhance the representation of key regional features. It adopts the efficient AKConv module to replace standard convolutions, improving multi-scale defect perception while reducing parameter overhead. Furthermore, a DySample dynamic sampling module is introduced in the upsampling stage to better reconstruct feature details and preserve semantic information. An Inner-GIoU regression loss function is designed to optimize bounding box localization accuracy.Experimental results on rail surface defect detection tasks demonstrate the effectiveness of the proposed model. EADI-YOLO achieves a mAP@0.5 of 90.3% and a mAP@0.5:0.95 of 49.1%, with an inference speed of 120 FPS, representing improvements of 2.8%, 6.7%, and 15 FPS over YOLOv10n, respectively. Compared to other mainstream detection models such as RT-DETR and YOLOv13n, EADI-YOLO demonstrates superior detection accuracy while maintaining excellent lightweight and real-time performance. Generalization experiments show that EADI-YOLO maintains an excellent balance between accuracy, speed, and model complexity across different rail defect datasets, making it suitable for intelligent detection of railway track defects.

Railway fasteners are critical components for maintaining track stability and ensuring the safe operation of trains. Accurate and real-time detection of fastener defects is essential for achieving intelligent railway maintenance. However, existing object detection algorithms often struggle to balance detection accuracy, computational efficiency. To address these challenges, this paper proposes an improved lightweight detection algorithm—PGCF-YOLO (Pyramid Split Attention - GSConv + VoV-GSCSP – CARAFE – FocalerIoU – MPDIoU – You Only Look Once), based on the YOLOv8n architecture, aimed at enhancing the overall performance of the model. First, a Pyramid Split Attention (PSA) module is integrated into the backbone to strengthen the model’s capability in perceiving complex defect features. Then, lightweight GSConv and VoV-GSCSP modules are introduced into the Neck to reduce parameter count and computational overhead while maintaining strong feature extraction capacity. The CARAFE upsampling operator is adopted to replace traditional nearest-neighbor interpolation, improving the model’s ability to capture both fine-grained details and global semantic information. Finally, a novel regression loss function, Focaler-MPDIoU, is proposed to enhance bounding box localization accuracy and accelerate training convergence. Experimental results demonstrate that PGCF-YOLO achieves excellent inference efficiency while maintaining high detection accuracy, reaching 99.3% mAP and 128 FPS, which represent improvements of 2.2% in accuracy and 12 FPS in speed over the original YOLOv8n. Furthermore, the parameter count and GFLOPs are reduced by 11.0% and 17.3%, respectively. Compared with other mainstream object detection models, PGCF-YOLO demonstrates superior performance in detection accuracy, inference speed, and model size. Experimental results on the E-Type fastener dataset demonstrate that PGCF-YOLO possesses strong generalization capability, validating its practical applicability in railway fastener defect detection tasks.