Ablation experiment of PE Unet for MPRS data set.

Image transfer plays a significant role in the manufacture of PCB; it affects the production speed and quality of the manufacturing process. This study proposes a surface-framework structure, which divides the network into two parts: surface and framework. The surface part does not include subsampling to extract the detailed features of the image, thereby improving the segmentation effect when the computing power requirement is not large. Meanwhile, a semantic segmentation method based on Unet and surface-framework structure, called pure efficient Unet (PE Unet), is proposed. A comparative experiment is conducted on our mark-point dataset (MPRS). The proposed model achieved good results in various metrics. The proposed network’s IoU attained 84.74%, which is 3.15% higher than Unet. The GFLOPs is 34.0 which shows that the network model balances performance and speed. Furthermore, comparative experiments on MPRS, CHASE_DB1, TCGA-LGG datasets for Surface-Framework structure are introduced, the IoU promotion clipped means on these datasets are 2.38%, 4.35% and 0.78% respectively. The Surface-Framework structure can weaken the gridding effect and improve the performance of semantic segmentation network.图像转印在印制电路板（Printed Circuit Board，PCB）的制造中发挥着关键作用，其直接影响制造流程的生产效率与产品质量。本研究提出一种表面-框架（surface-framework）结构，将网络划分为表面与框架两个分支：表面分支不采用下采样操作，以提取图像的细节特征，从而在计算资源需求较低的场景下提升分割效果。与此同时，本研究提出一种基于Unet与表面-框架结构的语义分割方法，命名为纯高效Unet（PE Unet）。研究团队在自建的标记点数据集（MPRS）上开展了对比实验，所提模型在多项评估指标上均取得了优异表现：该网络的交并比（Intersection over Union，IoU）达到84.74%，较原始Unet提升3.15%；其十亿次浮点运算每秒（GFLOPs）为34.0，表明该模型在性能与推理速度之间实现了良好平衡。此外，本研究针对表面-框架结构在MPRS、CHASE_DB1、TCGA-LGG三个数据集上开展了对比实验，上述数据集上的交并比提升截断均值分别为2.38%、4.35%与0.78%。结果表明，表面-框架结构能够有效削弱网格伪影，提升语义分割网络的整体性能。