计算机科学
算法
假阳性悖论
特征(语言学)
光学(聚焦)
瓶颈
人工智能
嵌入
特征提取
推论
GSM演进的增强数据速率
目标检测
模式识别(心理学)
计算机视觉
边缘检测
比例(比率)
功能(生物学)
班级(哲学)
算法设计
背景减法
曲面(拓扑)
作者
Shaobin Chen,Xinru He,Feifei Lian,Yani Lin,Y Huang,Jiawen Li,Fengyi Lin
出处
期刊:
[SPIE]
日期:2026-07-13
卷期号:: 6-6
摘要
To address the limitations of lightweight detection algorithms for rail surface defects—overemphasis on local features at the cost of global context, causing small-defect missed detection, high false positives and unsatisfactory real-timeaccuracy trade-offs—this paper proposes an enhanced YOLO-EOA algorithm based on global feature learning. Built on YOLOv12, the algorithm introduces the Enhanced Outlook Attention (EOA) module. Through multi-scale dynamic convolutions, coordinate attention fusion and lightweight bottleneck structures, it achieves global-local feature synergistic extraction for multi-dimensional precise defect detection. Embedding EOA into the backbone and neck networks strengthens spatial semantic correlation and feature propagation. Meanwhile, the loss function is optimized: Scale-Adaptive Focus Loss (SAFL) resolves class and scale imbalance, and directional constraints are integrated into CIoU loss to fit the elongated rail structure. Experimental results show that YOLO-EOA obtains 0.935 mAP, 0.916 precision and 0.853 recall. With only 2.9M parameters and 7.2 GFLOPs, it achieves an inference speed of 27 FPS on the Huawei Ascend Atlas 200DK edge device, which is also significantly better than mainstream models. It improves smalldefect detection rates and reduces background false positives while maintaining lightweight deployability. It satisfies the real-time requirements of vehicle-mounted inspection, providing an efficient solution for automated railway defect detection.
科研通智能强力驱动
Strongly Powered by AbleSci AI