【摘要】 为了在装配前对船舶零部件不规则形状表面缺陷进行高精度识别和实时定位。设计一种基于动态结构重参数化策略的轻量级深度神经检测网络，用于多尺度表面缺陷的高效检测。该动态学习策略通过在训练过程中扩展具有高重要性的骨干网结构提取更多信息。同时在边缘设备上进行检测时，可利用结构重参数化将模型无损压缩为轻量级模型实现高效推理，最大化利用物联网场景中边云不同的计算性能。通过实验验证本文所提出方法相比传统方法具有更好的优越性，为船舶装配零部件自动化高效检测提供有效支撑。更多还原

【Abstract】 High-precision identification of surface defects and real-time localization of irregular shapes on ship components before assembly are crucial. A lightweight detection algorithm using a dynamic structure reparameterization strategy is designed for efficient detection of multi-scale surface defects.Based on this dynamic learning strategy, more information is extracted by expanding the backbone network structure with high importance during training. Meanwhile, the structural reparameterization can be used to losslessly compress the model into a lightweight model for efficient inference when the detection is performed on edge devices. Thus, the computational performance of different edge clouds in IoT scenarios can be maximized. The method proposed has been proven to be better and superior than traditional methods by experiment results, providing effective support for the automated and efficient detection of parts for ship assembly.更多还原