【摘要】 为解决铝型材表面缺陷小目标和多尺度等问题，通常采用大规模网络来获得较好的检测结果，但会使实时性较差。针对上述问题，提出一种基于知识蒸馏的模型压缩技术，它将教师网络中特征知识迁移到小规模学生网络中，以实现检测的高速和高精度。为了更好地进行多尺度特征迁移，设计一种多尺度注意力特征蒸馏模块(MADM)。针对铝型材表面缺陷图像中小目标多和分类与定位任务的差异性，设计一种基于空间注意力机制的蒸馏模块(SADM)。实验结果表明，所提出模型的大小为20.07 MB,检测精度达到73.8%,实现检测速度和精度的均衡。进行实时检测，速度达到64帧每秒，满足实时需求。更多还原

【Abstract】 To address the challenge of detecting small targets and surface defects of aluminum profiles at multiple scales, large-scale networks are often employed to achieve better detection performance. However, this approach can hinder real-time performance. To overcome this limitation, this paper proposes a model compression technique based on knowledge distillation that transfers feature knowledge from a teacher network to a smaller student network, enabling high-speed and high-accuracy detection. We designed a multi-scale attention feature distillation module(MADM) to facilitate multi-scale feature transfer, as well as a spatial attention-based distillation module(SADM) to handle the numerous small targets and differences in classification and localization tasks present in aluminum surface defect images. Experimental results show that the proposed model has a size of 20.07 MB and mAP of 73.8%, achieving a balance of detection speed and accuracy. Additionally, real-time detection is performed in this study, with a speed of 64 FPS, meeting real-time requirements.更多还原