It is essential to perform more accurate classification of surface defects for hot-rolled strips, in order to guarantee manufacturing efficiency and improve product quality in steel industry. Surface defects’ complexity poses significant challenges for traditional machine vision methods, which struggle with detecting minute defects and the high degree of shape similarity between defect classes. This study introduces a novel methodology to improve classification accuracy by integrating transfer learning with the XGBoost algorithm. Specifically, a pre-trained VGG16 deep convolutional neural network is utilized as a feature extractor to capture high-level representations from steel surface images, while the XGBoost algorithm serves as the classifier to distinguish among six common types of surface defects. Experimental results indicate that the proposed VGG16-XGBoost model substantially enhances classification performance, attaining a notable accuracy of 98.05%. This approach illustrates the potential of combining deep feature extraction with XGBoost to address complex surface defect classification challenges in steel manufacturing.