Abstract: Steel surface defects have a significant impact on the quality and performance of many industrial products, which will bring huge economic losses to production. Therefore, it is very necessary to detect the steel surface in real time and find defects in time. In order to improve the detection performance of steel surface defects with large scale differences and complex backgrounds, a steel surface defect detection model based on multiscale local and global context information was proposed. Convolution operation with down-sampling mechanism in the convolutional neural network model was used to obtain rough multi-scale local feature maps. Then, self-attention mechanism was used to act on the local feature map extracted by convolution at each scale, to obtain long-distance dependencies between pixels (such as scratches, patches, inclusions, etc.), thus to enhance the inter-class discrimination ability of defects. Afterwards, the feature pyramid structure was used to fuse multi-scale feature maps, to improve the detection ability of multi-scale objects. Finally, channel and spatial attention module and WIoU loss function were introduced. The experimental results showed that, compared with algorithms such as Faster RCNN and EDDN, the proposed method was effective in improving the detection performance of steel surface defects.