Abstract: Aiming at the deficiencies of existing algorithms that suffer from insufficient global semantic feature extraction capability with large parameter sizes,to achieve semantic edge detection of bipolar plate welding area,a lightweight model for Edge Detection with Transformer (EDTER) termed Edge Detection with Mix Transformer (EDFormer) was developed.By incorporating a learnable Deformable Patch module into EDTER two-stage framework for generating edge prediction maps,EDFormer avoided the destruction of semantic information that could occur with fixed-size patches.Additionally,the columnar encoder was replaced with a hierarchical Mix Transformer structure and a lightweight decoder was designed based on Multi-layer Perception (MLP),which could enhance multi-scale feature extraction capabilities while maintaining a compact parameter size,fulfilling the deployment criteria for low-powered computational edge devices.Experimental results demonstrated that EDFormer outperformed EDTER on the public dataset BSDS500.On the self-built dataset,EDFormer achieved ODS,OIS,and AP values of 97.7%,98.1% and 98.3% respectively,while reducing the parameter scale by three-quarters,aligning with industrial application requirements.

Key words: bipolar plate welding, semantic edge detection, deep learning, Transformer

摘要： 针对现有双极板焊接区域语义边缘检测算法存在的全局语义特征提取能力不足与参数规模大的问题,基于Transformer边缘检测模型（EDTER）提出了轻量化的Mix Transformer边缘检测模型（EDFormer）。在EDTER“两阶段”生成边缘预测图的框架基础上,引入了可学习的Deformable Patch模块,避免了固定尺寸Patch对语义信息的破坏。将柱状的编码器替换为分层的Mix Transformer结构,并使用了轻量的MLP解码器,以更小的参数规模实现了更好的多尺度特征捕捉能力,满足了在弱算力边缘设备上的部署需求。实验结果表明,相较于EDTER,EDFormer在公共数据集BSDS500上性能稳定提升。在自建数据集上ODS、OIS及AP分别达到了97.7%、98.1%、98.3%,参数规模缩小了3/4,符合工业应用需求。

关键词: 双极板焊接, 语义边缘检测, 深度学习, Transformer模型