针对高铁部件外观裂痕检测中裂痕特征细微、裂痕样本稀缺导致的检测查全率低、误检率高及泛化能力不足等问题，提出一种基于改进YOLO与SimpleNet的双阶段外观裂痕检测算法（YOLO-SimpleNet-FD）。首先，在YOLOv8分割模型的骨干网络中引入C2f＿Faster＿EMA模块，提升部件提取的精度与效率；其次，在SimpleNet中集成Ghost Bottleneck模块增强特征提取能力，并引入混合局部通道注意力（MLCA）机制增强对细微裂痕特征的关注；最后，先利用YOLOv8精准分割部件区域，有效剔除背景干扰，再通过改进SimpleNet实现裂痕识别与定位。试验结果表明，所提算法的部件提取交并比（IoU）、查全率（Recall）、误检率（FPR）、推理速度（FPS）分别达到92.3%、95.6%、2.1%、52 FPS；相较于原始融合算法，IoU提升4.1百分点，查全率提升4.4百分点，误检率提升2.2百分点，推理速度基本保持稳定；相较于单一SimpleNet算法，查全率提升16.7百分点，误检率降低10.9百分点。该算法为高铁部件外观裂痕检测提供了高效、可靠的技术方案，具有良好的工程应用价值。

Aiming at the problems of low detection recall rate, high false detection rate and insufficient generalization ability caused by the subtle crack features and scarce crack samples in the appearance crack detection of high-speed rail components, a two-stage appearance crack detection algorithm based on improved YOLO and SimpleNet(YOLO-SimpleNet-FD) is proposed. Firstly, the C2f＿Faster＿EMA module is introduced into the backbone network of the YOLOv8 segmentation model to improve the accuracy and efficiency of component extraction. Secondly, the Ghost Bottleneck module is integrated into SimpleNet to enhance the feature extraction ability, and the Mixed Local Channel Attention(MLCA) mechanism is introduced to enhance the attention to subtle crack features. Finally, the YOLOv8 is used to precisely segment the component area to effectively suppress background interference, and then the improved SimpleNet is used to achieve crack recognition and localization. The experimental results show that the intersection over union(IoU), recall rate, false detection rate and inference speed of the proposed algorithm for component extraction reach 92.3%, 95.6%, 2.1%and 52 FPS, respectively. Compared with the original fusion algorithm, the IoU is decreased by 4.1 percentage points, the recall rate is increased by 4.4 percentage points, the false detection rate is increased by 2.2 percentage points, and the inference speed remains basically stable. Compared with the single SimpleNet algorithm, the recall rate is increased by 16.7 percentage points and the false detection rate is reduced by 10.9 percentage points. This algorithm provides an efficient and reliable technical solution for the appearance crack detection of high-speed rail components and has good engineering application value.