【摘要】 针对电子接插件生产过程中表面焊点的漏焊、过焊，裂纹等外观缺陷问题，提出一种开集分类Openmax方法优化网络节点MoblieNetV2模型完成焊接件外观缺陷的自动识别检测方法。将算力要求较高的模型训练过程在高性能PC端完成，训练完成的优化识别模型部署至资源受限的嵌入式终端完成工厂环境的图像识别推理检测。详细阐述了小样本焊接件标准与缺陷图像下的数据集的采样，增强与标记，PC端网络模型的构建、优化与训练，以及嵌入式终端推理模型的工程组织设计与推理实现。工厂实测表明，所设计的网络模型标准焊件识别率达到95.5%,缺陷件分类识别率为96%,符合厂家的识别率要求。更多还原

【Abstract】 An Openmax method is proposed to optimize the network node MoblieNetV2 model for automatic identification and detection of surface defects such as solder leakage, over soldering, and cracks in the production process of electronic connectors. The training process of models with high computing power requirements is completed on high-performance PCs, and the optimized recognition model trained is deployed to resource constrained embedded terminals to complete image recognition inference and detection in factory environments. The sampling, enhancement, and labeling of datasets under small sample welding standards and defect images, and the construction, optimization, and training of PC network models, as well as the engineering organization design and inference implementation of embedded terminal inference models are elaborated. Factory tests show that the network model designed has a standard welding recognition rate of 95.5% and a defect classification recognition rate of 96%,which meets the manufacturer’s recognition rate requirements.更多还原