Abstract: Surface defect detection is an important step in controlling the quality of tungsten rods. Due to the complex background, high noise interference, low contrast, low defect resolution, and wide scale span of tungsten rod images, the detection accuracy of the existing YOLOv5 algorithm is difficult to meet the requirements of industrial sites. This paper proposed an improved algorithm based on YOLOv5, which first reconstructed the detection head of the network model and used more scale feature maps for result prediction; Secondly, an attention mechanism was added to the feature extraction network to enhance the algorithm's ability to extract key features. At the same time, a weighted bidirectional feature pyramid network structure was adopted to achieve multi-scale feature fusion, reducing redundancy in the feature fusion process and improving the fusion degree of multi-scale target features. The experimental results showed that the tungsten rod surface defect detection algorithm based on improved YOLOv5 was more suitable for metal surface defect detection problems with small size and large curvature characteristics, and could significantly improve the accuracy and precision of such detection.