In the task of motherboard defect detection, the high similarity between defect targets and the motherboard background, coupled with the small proportion of defect targets, lead to a high likelihood of defects being missed or misdetected. To address this issue, a motherboard defect detection method based on context information fusion and dynamic sampling, named BC-YOLO, was proposed. First, a context fusion attention module, named MHCA, was constructed. Contextual information was effectively fused by independently generating feature maps and utilizing an attention mechanism. This enhanced the detection performance of easily confusable targets. Second, a dynamic upsampling module, named FMDU-Upsample, was designed. Multi-scale feature extraction was combined with dynamic sampling methods to improve the quality of feature map upsampling. Sampling positions were adaptively adjusted according to input features by utilizing dynamic offsets. Multi-scale feature extraction was incorporated to enhance the representational capability of feature maps. Finally, SPD-conv was introduced into the feature extraction network. Information in the tensor was reorganized to enable the network to better handle features at different scales, thus improving the sensitivity of convolution operations to local information. Experimental results indicated that the BC-YOLO algorithm achieved a mean average precision (mAP) of 98.6% on the motherboard defect dataset, representing an improvement of 2.4 percentage points over the previous YOLOv9 algorithm. The average detection speed of the YOLOv9 model was 52.6 frames per second, outperforming other advanced models. Experiments on the public dataset GC10-DET demonstrated that the proposed method had strong generalization ability. The detection accuracy was improved by 3.5 percentage points compared with the previous YOLOv9 algorithm, allowing for more precise identification of defect targets.