Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (16): 269-275.DOI: 10.3778/j.issn.1002-8331.2306-0384

• Graphics and Image Processing • Previous Articles     Next Articles

DeepLabv3+ Lightweight Image Segmentation Algorithm Based on Multilevel Feature Fusion

ZHOU Huaping, DENG Bin   

  1. School of Computer Science and Engineering, Anhui University of Science and Technology, Huainan, Anhui 232001, China
  • Online:2024-08-15 Published:2024-08-15

融合多层次特征的DeepLabv3+轻量级图像分割算法

周华平,邓彬   

  1. 安徽理工大学 计算机科学与工程学院,安徽 淮南 232001

Abstract: Image semantic segmentation models that rely on deep learning usually possess a vast number of parameters, exhibit high complexity, and pose challenges when deploying on mobile platforms. To address the aforementioned issues, this paper enhances the DeepLabv3+ algorithm and introduces an improved lightweight image segmentation algorithm. Firstly, the backbone network of the model uses the lightweight MoblieNetv2 network, and obtains four different levels of input characteristics and four different semantic information. The CAFF (coordinate attention feature fusion) module is proposed, which integrates the features of the middle two levels and adds location information. The ASPP (atrous spatial pyramid pooling) module is improves, the CS_ASPP module is proposed. The CAM (channel attention module) mechanism is introduced after convolution with different expansion rates, while parallel strip pooling (SP) is used to obtain contextual information, and SAM (spatial attention module) mechanism is introduced after feature fusion to improve segmentation accuracy. In the experiment conducted on the PASCAL VOC 2012 dataset, the mIoU (mean intersection over union ratio) is measured to be 79.14%. Compared with common segmentation algorithms and improved segmentation algorithms, the model demonstrates the improved accuracy and achieves a desirable equilibrium among the number of parameters, segmentation speed, and segmentation performance.

Key words: image segmentation, DeepLabV3+, multilevel feature fusion, lightweight, attention mechanism

摘要: 基于深度学习的图像语义分割模型通常参数量大,复杂度高,难以部署到移动平台。针对以上问题,对DeepLabv3+算法进行改进,提出一种改进的轻量级图像分割算法。模型的骨干网络使用轻量级MoblieNetv2网络,并获取四个不同层次的输入特征,得到四种不同的语义信息;提出CAFF(coordinate attention feature fusion)模块,融合中间两个层次特征并加入位置信息;改进空洞空间金字塔池化(atrous spatial pyramid pooling,ASPP)模块,提出CS_ASPP(channel strip_atrous spatial pyramid pooling)模块,在不同膨胀率的空洞卷积后引入CAM(channel attention module)机制,同时并联带状池化(strip pooling,SP)获取上下文信息,并在特征融合后引入SAM(spatial attention module)机制提升分割精度。在PASCAL VOC 2012数据集上进行实验,平均交并比(mIoU)达到了79.14%。实验结果表明,该模型更加精准,且在参数量、分割速度以及分割效果之间达到了较好的平衡。

关键词: 图像分割, DeepLabV3+, 多层次特征融合, 轻量级, 注意力机制