Scholarlink

<p>In order to solve the problems of large target size difference and complex background in the portrait segmentation task, it is easy to misjudge small targets, and the contour segmentation of people and joint parts such as elbow is not fine enough. A portrait segmentation model based on semantic injection and multi-scale cross-axis attention mechanism is proposed. First, the MobileNetV2 module is used to generate local semantic information quickly and reduce the model computational overhead; second, for the reasons of large difference in the size of the person and the complex background, in order to capture the multi-scale feature information and establish the long-distance dependency between pixels more effectively. However, due to the relative complexity of the Transformer and the high demand for computational resources, this paper uses the multi-scale cross-axis attention mechanism to capture the long-distance dependencies and enhance the feature representation. Meanwhile, in order to avoid the problem of character and context misclassification, we use a semantic injection module, which is used to reduce the discrepancy between the shallow semantic information and the deep semantic information. The results on PenFudanPed and CIHP datasets show the best experimental metrics compared to other models, with mIoU values improved by 1.3% and 1.52% over the optimal model, respectively. It also has the best segmentation effect for visualization of predicted results for features of different scales and poses. In addition, we conduct experiments on the ADE20K dataset to further validate the robustness and generalisation of the model.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Semantic injection and multi-scale cross-axis attention-based portrait segmentation model

  • 期刊论文

  • 发表日期：

    2025年3月10日

查看全文
  • Yan Cheng,
  • Yanying Chen,
  • Zou Huang,
  • Guowei Li,
  • Gongcheng Yu,
  • Jiaqi Wang,
  • Yuxiao Yao

摘要

In order to solve the problems of large target size difference and complex background in the portrait segmentation task, it is easy to misjudge small targets, and the contour segmentation of people and joint parts such as elbow is not fine enough. A portrait segmentation model based on semantic injection and multi-scale cross-axis attention mechanism is proposed. First, the MobileNetV2 module is used to generate local semantic information quickly and reduce the model computational overhead; second, for the reasons of large difference in the size of the person and the complex background, in order to capture the multi-scale feature information and establish the long-distance dependency between pixels more effectively. However, due to the relative complexity of the Transformer and the high demand for computational resources, this paper uses the multi-scale cross-axis attention mechanism to capture the long-distance dependencies and enhance the feature representation. Meanwhile, in order to avoid the problem of character and context misclassification, we use a semantic injection module, which is used to reduce the discrepancy between the shallow semantic information and the deep semantic information. The results on PenFudanPed and CIHP datasets show the best experimental metrics compared to other models, with mIoU values improved by 1.3% and 1.52% over the optimal model, respectively. It also has the best segmentation effect for visualization of predicted results for features of different scales and poses. In addition, we conduct experiments on the ADE20K dataset to further validate the robustness and generalisation of the model.