A surface defect semantic segmentation algorithm based on improved ResUnet is proposed to address the issues of low detection rate and large differences in detection shapes caused by the diversity of size and shape of metal surface defects. This method is based on the ResUnet network structure and uses advanced ConvNeXt-T to improve the backbone network and extract more representative features; Design and add a global context module during the decoding phase to enhance the model"s ability to model global context; At the same time, the fusion of multi-level and multi-scale features in the decoding stage makes the model more suitable for steel surface defects with variable sizes and sizes, improving the accuracy of defect recognition. The effectiveness of the proposed defect recognition algorithm was quantitatively and qualitatively validated on the steel surface defect dataset provided by Xavier Steel Company. Compared with the comparison method, the value of the proposed method reached the highest of 0.7784, and the running time of a single image on the GPU was only 14.7ms. At the same time, the subjective visual effect of the defect segmentation results of the proposed method was closest to the label, indicating that the method has good robustness, accuracy, and efficiency.