Abstract: In response to the low efficiency of existing magnetic levitation train foreign object intrusion detection systems and the problems of misdetection and missed detection of foreign objects of different sizes, a foreign object detection algorithm based on improved YOLOv8n, named YOLO-RFCL (robust feature calibration lightweight network), is proposed. Adopting the robust feature downsampling (RFD) module to integrate various downsampling features, the model’s feature expression ability is optimized. It introduces frequency adaptive dilation convolution (FADC), uses a modulation mechanism to assign different dilation rates for each pixel, reduces the model’s complexity and computational cost. It establishes a context feature calibration network (CFCN) to improve small target detection accuracy through self-calibration and local detail fusion. It proposes the DEHead (detail-enhanced head) to enhance the model’s representation and generalization capabilities. Experimental results show that on the custom magnetic levitation train foreign object intrusion dataset, YOLO-RFCL improves the precision (P), recall (R), mAP50, and mAP95 by 2.3, 6.6, 4.3, and 2.5 percentage points, respectively, compared to the original algorithm, and reduces the parameter count (Params) by 3.7%. Compared to YOLOv11n, YOLO-RFCL improves precision (P), mAP50, and mAP95 by 3.9, 2.9, and 2.4 percentage points, respectively. The detection speed of YOLO-RFCL reaches 111.8 FPS, validating the effectiveness and accuracy of the algorithm.

Key words: foreign object intrusion detection, YOLOv8n, adaptive extended convolution, feature calibration, feature fusion

摘要： 针对现有磁浮列车异物入侵检测效率低以及对不同尺寸异物存在错检、漏检等问题，提出一种基于改进YOLOv8n的异物检测算法YOLO-RFCL（robust feature calibration lightweight network）。采用鲁棒特征下采样模块（robust feature downsampling，RFD），融合多种下采样特征，优化模型特征表达能力。引入频率自适应扩张卷积（frequency adaptive dilation convolution，FADC），采用调制机制为每个像素分配不同的膨胀率，降低模型的复杂度和计算成本。建立上下文特征校准网络（context feature calibration network，CFCN），通过自校准和局部细节融合，提升小目标的检测精度。提出DEHead检测头（detail-enhanced head），增强模型的表征和泛化能力。实验结果表明，在磁浮列车异物入侵自定义数据集上，YOLO-RFCL与原算法相比精确率P、召回率R、mAP50、mAP95分别提升了2.3、6.6、4.3、2.5个百分点，参数量Params减少了3.7%，YOLO-RFCL与YOLOv11n相比精确率P、mAP50、mAP95分别提升了3.9、2.9、2.4个百分点；YOLO-RFCL的检测速度达111.8 FPS，验证了算法的有效性和精确性。

关键词: 异物入侵检测, YOLOv8n, 自适应扩张卷积, 特征校准, 特征融合