<p>Detection of weld proximity defects aims to recognize and classify different proximity defects in industrial welding. It presents some characteristics such as the small-target, localized diffusion, and dense distribution, which is difficult in steel-sheet welding engineering. Due to small-target size, localized diffusion, the similar and dense distribution, and imbalanced sample numbers, the existing detection models with the fixed receptive field are weak in detecting weld proximity defects. To solve the problem, we propose an expansive you-only-look-once detection model, called DRB-YOLOv5. It integrates a novel dilation causal convolution cross-stage partial network and the improved Focal Loss function. First, we provide a dilation causal transposed-convolution residual block cross-stage partial network to amplify and extract features of weld proximity defects effectively. Further, we reduce the block number of two middle Resblock Bodies in the Backbone of DRB-YOLOv5 to avoid network overfitting. Second, we design a Prediction Head via the Bias Focal Loss to balance positive and negative samples and improve the detection accuracy of various number-imbalanced weld defects. Finally, we compare the proposed model with some related existing models via the real weld proximity defects dataset. Experimental results show that the proposed DRB-YOLOv5 can present superior effects, mainly including mAP: 89.21%, F1: 83.50%, and recall: 80.73%.</p>