Surface defect detection of steel strips is crucial for ensuring product quality in steel manufacturing. The achievement of efficient and accurate detection is significant for product performance. While deep learning methods have made significant progress in defect detection, two challenges persist in practical applications: First, due to the pursuit of high yield rates in industrial production, defect samples are limited and sample annotation is time-consuming. Secondly, the collected samples may contain redundant features, affecting model training efficiency and generalization performance. To address feature redundancy, a dynamic sample selection method based on feature representation and learning feedback is proposed. Initially, a multi-dimensional feature quantification model incorporating geometric morphology, grayscale distribution, and directional features is formulated to characterize defect features. Subsequently, a feature representationbased sample selection strategy is designed to select diverse and representative training samples through feature clustering. Finally, a confidence-based dynamic optimization strategy is proposed to obtain supplementary samples through learning feedback, achieving adaptive optimization. Experimental results on the NEU-DET dataset show that the method achieves a mean average precision of 76.99% while reducing the training sample size by 52%. It is comparable with using the complete dataset and decreases training iteration time by 62%. Evaluation of various detection models shows that the method effectively improves efficiency while maintaining performance across different architectures.