With the widespread application of deep learning in industrial surface defect detection, YOLO-series object detection models have become mainstream due to their compact structure and high inference efficiency. However, deployment on embedded devices is limited by computational resources. Therefore, effective model compression is required for lightweight deployment. In this study, a joint compression strategy combining channel pruning and feature distillation is proposed for the YOLOv7-Tiny model. Specifically, the Group_slim method is first applied to perform channel-level structural compression, which removes redundant channels and reduces model complexity. Then, the original uncompressed model is used as the teacher network. The student network is guided using the Channel-Wise Distillation (CWD) method to learn key feature representations after pruning. Extensive experiments on the MVTecAD-Screw industrial defect dataset show that the proposed strategy reduces model parameters and computational cost while maintaining high accuracy. A good balance between inference speed and precision is achieved. Compared with other single pruning or knowledge distillation methods, the Group_slim + CWD scheme demonstrates superior overall performance across multiple evaluation metrics.