In view of the defects of steel and aluminum products in industrial production lines, a modified method for steel surface defect detection was proposed for Faster R-CNN, with a view to reducing labor costs and improving the efficiency and quality of production. Based on the normalized attention mechanism, the modified model weights the feature information of the feature target to capture the correlation of local feature information, so that more attention is paid to the feature information of target, thereby improving the detection precision. Furthermore, multi-level feature fusion module Feature Pyramid Network (FPN) is used for further multi-level fusion of features output from the backbone network, which enhances the performance of small features in the deep-level network structure and addresses disappearance of small features due to deepening of neural networks. Experiments have proved the feasibility of this method on the public data set (“Aluminum profile defect identification” provided by the Guangdong Innovation Competition on Industrial Intelligent Manufacturing Big Data). Compared with Faster R-CNN, SSD, YOLOv4, YOLOv5 and YOLOv7, the proposed modified Faster R-CNN model is more precise.