现阶段,图像深度学习算法无法检测时序性的工艺流程问题。 本文针对针织机械山板总成的人为装配工艺进行研究, 提出 MS-RetinaNet 目标检测算法。 借鉴自然语言处理的思想,引入 Swin-Transformer 结构,保留了 CNN 结构的层次性,弥补了 CNN 结构对于高层语义信息融合不足的问题,增强了全局与细节学习能力;使用改进的 GIoU Loss,增加判定因子式,缓解损失 计算退化的影响,优化边界框回归效果;根据多尺度目标参数,采用最佳锚框比,提高了召回率和检测精度;设计时序检测头,使 算法具有判别目标先后顺序和逻辑关系的能力。 实验结果表明,算法 AP 可达 90. 3%,高于当前主流算法 2%以上,单张图片检 测速度约 46 ms,满足了工艺流程的时序检测要求,综合性能优越。
At this stage, the image deep learning algorithm cannot detect the chronological process problem. In this paper, the artificial assembly process of the mountain board assembly of knitting machinery is studied, and the MS-RetinaNet object detection algorithm is proposed. Using the idea of natural language processing for reference, the Swing-Transformer structure is introduced to retain the hierarchy of CNN structure, make up for the lack of high-level semantic information fusion in CNN structure, and enhance the ability to learn overall and details. The improved GIoU Loss is used to increase the judgment factor formula, mitigate the impact of loss calculation degradation, and optimize the regression effect of the bounding box. According to the multi-scale target parameters, the best anchor frame ratio is adopted to improve the recall rate and detection accuracy. The chronological detector is designed to enable the algorithm to distinguish the sequence and logical relationship of the target. The experimental results show that the algorithm AP can reach 90. 3%, which is more than 2% higher than the current mainstream algorithm. The detection speed of a single image is about 46 ms, meeting the chronological detection requirements of the process flow, and the overall performance is superior.
李 玮,高 林.改进 RetinaNet 的工艺流程检测算法[J].电子测量与仪器学报,2023,37(7):104-112复制