Abstract:

As an important modern electronic component, component segmentation of highly integrated Printed Circuit Boards (PCB) is a typical small-object image segmentation task. Owing to the excessive introduction of complex backgrounds, component segmentation of PCB images faces challenges such as insufficient boundary-awareness capability. To improve the boundary-awareness capability, an externally coordinated nested U-shape network structure called U²ECNet is proposed. Specifically, the backbone network of the algorithm is a nested U-shape structure, and an external expansion module is used in the coding and decoding system to efficiently learn global and local information and focus on the edge and corner details in the component region. A bootstrap refinement module is used to optimize the segmentation accuracy of the model and improve the segmentation effect of PCB components by aggregating the global semantic information through multi-scale feature mapping. PCB_SOD, a new image segmentation dataset containing 5 608 training images and 2 403 test images, was used to perform the segmentation task and trained in the proposed network. Experiments on the DUTS and PCB_SOD datasets show that the network model achieved a Mean Absolute Error (MAE) of 0.045 and 0.027 and max F_β of 86.1% and 87.2%, respectively, demonstrating reduced MAE and improved max F_β, achieving the best overall segmentation performance compared to other methods. The proposed externally coordinated nested U-shape structure improves the accuracy of PCB component segmentation, exhibits good robustness in complex backgrounds, and generates the most accurate saliency segmentation graph.

Key words: significance segmentation, U-shape network structure, PCB_SOD dataset, boundary perception, guiding refinement

摘要：

作为重要的现代电子元器件, 高度集成的印刷电路板(PCB)的组件分割属于典型的小物件图像分割。由于复杂背景的过度引入, PCB图像的组件分割面临着边界感知能力不足等挑战。为了提高PCB分割组件边界感知能力, 提出一种外部协调的嵌套U-型网络结构(U²ECNet)。U²ECNet的主干网络为嵌套U-型结构, 在编解码体系中使用外部扩张模块, 有效学习全局和局部信息, 并关注组件区域中的边缘和角细节; 使用引导细化模块, 通过多尺度特征映射聚合全局语义信息, 从而优化模型的分割精度, 同时提高对PCB组件分割的效果; 制作新的图像分割数据集PCB_SOD, 其包含5 608张训练图像和2 403张测试图像, 用于执行分割任务, 并在所提网络中进行训练。在DUTS和PCB_SOD数据集上的实验结果表明, U²ECNet在平均绝对误差(MAE)和maxF_β上分别达到0.045、86.1%和0.027、87.2%, 相较于其他方法, U²ECNet的MAE更低, maxF_β更高, 整体分割性能达到最佳。此外, 所提外部协调的嵌套U-型结构提升了PCB组件分割的精度, 在复杂背景中表现出良好的鲁棒性, 生成了准确的显著性分割图。

关键词: 显著性分割, U-型网络结构, PCB_SOD数据集, 边界感知, 引导细化