In view of the complex distribution of component positions， lack of prominent defect targets， and multi-scale issues in the detection of defects in computer motherboard assembly， this paper proposed an end-to-end defect detection algorithm based on parallel feature extraction and cross-attention progressive feature fusion. Firstly， a parallel residual feature extraction network was proposed by combining partial convolution and visual Transformer. The low computational complexity of partial convolution was utilized to extract local features， while the long-distance modeling ability of visual Transformer was utilized to expand the receptive field of the model and enhance the feature extraction ability of the network. Secondly， the cross-attention mechanism was introduced to progressively fuse multi-scale features， and a multi-scale cross-attention progressive feature fusion network was constructed to enhance the feature fusion ability of the detection model. The experimental results on the public dataset show that the mean average accuracy （mAP） of the algorithm reaches 94.63%， which is 4.62% higher than the baseline model YOLOv5 and is superior to several other advanced models. The detection speed reaches 25 FPS， achieving a good balance between detection accuracy and speed. It provides a fast and effective method for the automation and intelligence of surface assembly defect detection on computer motherboards in the actual industrial environment.