| 摘要: |
| 目的 考虑现有的图像语义分割网络存在分割精度低、参数量大等问题,提出一种融合轻量化和注意力机制
的语义分割算法。 方法 该算法在 DeeplabV3+网络模型结构的基础上,使用 MobileNetV2 网络替换原始网络模型结
构的 Xception 主干网络,构建轻量化语义分割网络结构,以此减少模型参数量和计算量,提高分割速度。 同时,为
了有效获取关注语义信息的正确特征,在编码阶段加入注意力模块机制,使网络模型在学习过程中只关注它所需
要关注的点,提高图像分割精度,达到良好的分割效果。 最后,在网络模型训练过程中引入 BCE loss( Binary Cross
Entropy loss) 和 Dice loss 损失函数相结合,加快网络的快速收敛,对模型更好的优化,以此提高模型的分割精度。
结果 通过在数据集 PASCAL VOC2012 实验验证表明,该算法的分割精度提高了 2. 82 个百分点,参数量降低了
14. 46 M。 同时,数据集 Cityscapes 的实验结果也验证了该算法的优越性。 结论 优化后的 DeeplabV3+网络模型提
高网络模型性能。 |
| 关键词: 语义分割 DeeplabV3+ 轻量化网络 通道注意力机制 损失函数 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| A Semantic Segmentation Algorithm Integrating Lightweight and Attention Mechanisms |
|
YUAN Manman1 LU Hao2
|
|
1. School of Computer Science and Engineering Anhui University of Science and Technology Anhui Huainan 232001
China
2. Anhui Civil Air Defense Architectural Design and Research Institute Hefei 230022 China
|
| Abstract: |
| Objective Considering the problems of low segmentation accuracy and a large number of parameters in existing
semantic image segmentation networks a semantic segmentation algorithm integrating lightweight and attention mechanisms
was proposed. Methods The algorithm replaced the Xception backbone network of the original network model structure
with the MobileNetV2 network based on the DeeplabV3+ network model structure and constructed a lightweight semantic
segmentation network structure so as to reduce the number of model parameters and computational volume and improve
the segmentation speed. Additionally an attention module mechanism was introduced during the encoding stage to
effectively capture correct features of focused semantic information enabling the network to focus only on relevant points
during the learning process thereby enhancing image segmentation accuracy and achieving satisfactory segmentation
results. Finally BCE loss binary cross entropy loss and Dice loss functions were combined in the network model
training process to accelerate the rapid convergence of the network and better optimize the model so as to improve the segmentation accuracy of the model. Results The experimental verification on the PASCAL VOC2012 dataset showed that
the segmentation accuracy of the algorithm was increased by 2. 82% and the number of parameters was reduced by 14. 46 M.
Furthermore experimental results on the Cityscapes dataset confirmed the superiority of the proposed algorithm.
Conclusion The optimized DeeplabV3+ network model enhances the performance of the network model. |
| Key words: semantic segmentation DeeplabV3+ lightweight network channel attention mechanism loss function |
