| 摘要: |
| 道路场景分割存在难以区分不同大小目标的问题.在PIDNet的启发下，本文提出一个新型的三分支网络，构建了大核卷积支路、Transformer支路和边界监督支路.通过边界监督支路改善了模型在不同大小目标上的分割效果；引入大核卷积模块，增加了模型的感受野；引入增强轴向注意力机制捕获轴向特征的长距离依赖，选择并融合Transformer支路与大核卷积支路的特征，最终融合三条支路的信息.将所提出的网络模型与PIDNet模型进行对比，其平均交并比（mIoU）值在CamVid数据集上提升了4.8个百分点，在Cityscapes数据集上提升了0.2个百分点，图像分割精度也有所提升，验证了混合模型在道路场景图像分割任务中具有优势. |
| 关键词: 图像分割 大核卷积 轴向注意力 边界监督 |
| DOI：10.20192/j.cnki.JSHNU（NS）.2025.02.001 |
| 分类号:TP399 |
| 基金项目: |
|
| A large kernel convolution hybrid axial attention semantic segmentation network based on boundary supervision |
|
XUE Yu, ZHANG Xiangfen, YUAN Feiniu
|
|
College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai 201418, China
|
| Abstract: |
| It was difficult to distinguish between targets of different sizes in road scene segmentation. In order to solve the above problem， inspired by PIDNet， a new type of three-branch network was proposed， which was constructed by a large kernel convolution branch， a Transformer branch and a boundary supervision branch. The segmentation effect of the model on different size targets was improved by boundary supervision branches， and the receptive field of the model was increased by introducing the large kernel convolution module. In the model， the enhanced axial attention mechanism was introduced to capture the long-distance dependence of axial features， and the features of the Transformer branch and the large kernel convolution branch were selected and integrated， and finally the information of the three branches was integrated. Compared with PIDNet， the mean intersection over union （mIoU） value of the proposed network model was increased by 4.8 percentage points on the CamVid dataset and 0.2 percentage points on the Cityscapes dataset respectively， and the image segmentation accuracy was also improved， which verified that the hybrid model had advantages in image segmentation task of road scenes. |
| Key words: image segmentation large kernel convolution axial attention boundary supervision |
