针对轮胎X射线图像灰度值较低、目标与背景之间对比度小、缺陷形状狭长以及缺陷目标较小导致的轮胎缺陷难检测的问题，提出了一种基于CSPDarknet53骨干网络和PANet（path aggregation network）颈部网络的轮胎X射线图像缺陷检测算法：TireX-DefectNet算法。首先，为了减少特征在网络传递过程中的丢失，在骨干网络中引入了可变形卷积和跨层次特征融合模块（cross-layer feature fusion, CLFF）；其次，为了提高网络对狭长目标的特征提取能力，在空间金字塔池化基础上构造了ASPPF（atrous spatial pyramid pooling with folding）模块，并将其引入到骨干网络中；最后，为了进一步提高检测精度，构造了CASA（channel attention and spatial attention）模块，使模型更加关注小目标缺陷的检测。实验结果表明，改进后的TireX-DefectNet算法较于其基线网络，精度提升了8.8%，为基于轮胎X射线图像的缺陷检测提供了一种有效方法。

A tire X-ray images defect detection algorithm based on CSPDarknet53 backbone network and PANet (path aggregation network) neck network called TireX DefectNet algorithm was proposed, which could improve the detection accuracy of tire X images caused by factors such as low grayscale values in photos, low contrast between targets and backgrounds, narrow defect shapes, and small defect targets. Firstly, in order to reduce the loss of features during network transmission, deformable convolution and cross layer feature fusion modules (CLEF) were introduced in the backbone network. Secondly, the ASPPF (atrous spatial pyramid pooling with folding) module was constructed and add to the backbone network, which was based on the spatial pyramid pooling module, and could improve the feature extraction ability of the network for narrow targets. Finally, the CASA (channel attention and spatial attention) module was constructed to make the model more focused on detecting small target defects, thereby further improving detection accuracy. The experimental results show that the improved TireX-DefectNet algorithm has improved accuracy by 8.8% compared to its baseline network, and provides an effective method for tire X-ray images defect detection.