Abstract: In order to solve the difficulty of defect location and classification in steel surface defect detection and improve the efficiency of steel surface defect detection, a steel surface defect detection algorithm based on multi-scale feature fusion and attention mechanism is proposed. The model of the algorithm is based on encoder-decoder structure to achieve the task of defect classification and segmentation. The encoder structure uses the residual network ResNet50 as the backbone and then uses the multi-scale feature fusion module to capture rich multi-scale spatial information. The decoder structure is based on the Global Attention Upsample module, and uses the global context weights generated by high-level semantic information to guide the shallow details to achieve more accurate selection of detailed information. Finally, the segmentation results are refined through 3×3 convolutional blocks, and gradually recover defect information and make predictions. Using the steel surface defect dataset provided by the kaggle competition platform to experiment with the algorithm, the Dice coefficient of defect detection can reach 94.22%. Compared with semantic segmentation models such as U-net, the defect detection effect is better.