Current object detection networks suffer from low accuracy and slow speed for industrial defect detection tasks. Industrial defect detection tasks are characterized by small area and large aspect ratio of the detected objects, as well as high speed requirements. We provide a label assignment strategy for defect shape characteristics to improve the training efficiency of a one-stage target detection network for defect detection scenarios. Also, label assignment distillation learning is used to obtain a model that takes into account the speed of detection. In this paper, experiments are conducted on several industrial defect datasets, and metrics such as mAP (mean average precision) values and inference speed are calculated. Compared with other models, the label assignment algorithm results in a 3% improvement in detection accuracy and a 58% acceleration in model inference speed after lightweighting.