基于U-Net网络的柱面透镜视觉定位策略

针对柱面透镜定位在摆盘工艺过程中产品良率低、生产作业时间长的问题,研究了一种基于U-Net网络的柱面透镜视觉定位策略。该策略结合柱面透镜摆盘工艺及特点,开发了柱面透镜视觉摆盘系统。将Blob分析和模板匹配作为前期视觉定位方法,通过高精度匹配获取工件点位信息以生成分割图对采集的原图进行标注,选取损失值最优的学习率对U-Net网络进行模型训练,以实现实际工件的分割定位。经过U-Net网络与高精度模板匹配针对不同型号的工件定位对比发现,相较于高精度模板匹配,U-Net网络可实现亚像素级别的定位误差。实践应用结果表明,该柱面透镜视觉摆盘系统视觉定位精度高,具有较强的产品适应能力,产品偏心合格率可达96%以上,满足工业摆盘任务要求。

Visual Positioning Strategy of Cylindrical Lens Based on U-Net Network

Aiming at the problems of low product yield and long production operation time during the plating processes of cylindrical lens positioning, a visual positioning correction strategy of cylindrical lens was studied based on U-Net network. The cylindrical lens plating processes and characteristics were combined in the strategy to develop a cylindrical lens visual plating system. The Blob analysis and template matching were used as the early visual positioning method, and the workpiece point information was obtained through high-precision matching to generate a segmentation diagram to annotate the original map of the acquisition, and the optimal learning rate of the loss values was selected to train the U-Net network to achieve the segmentation and positioning of the actual workpiece. U-Net was compared to high-precision template matching for different models of workpiece positioning. It is found that U-Net may achieve sub-pixel level positioning error relative to high-precision template matching. Practical application results show that the visual positioning accuracy of the cylindrical lens visual plating system has high accuracy and strong product adaptability, and the eccentric pass rate of the product may reach more than 96%, which meets the requirements of industrial plating tasks.