计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (24): 137-143.DOI: 10.3778/j.issn.1002-8331.2006-0437

• 模式识别与人工智能 • 上一篇    下一篇

基于DeepLab V3模型的图像语义分割速度优化研究

司海飞,史震,胡兴柳,杨春萍   

  1. 1.金陵科技学院 智能科学与控制工程学院,南京 211169
    2.哈尔滨工程大学 智能科学与工程学院,哈尔滨 150001
  • 出版日期:2020-12-15 发布日期:2020-12-15

Research on Speed Optimization of Image Semantic Segmentation Based on Deeplab V3 Model

SI Haifei, SHI Zhen, HU Xingliu, YANG Chunping   

  1. 1.College of Intelligent Science and Control Engineering, Jinling Institute of Technology, Nanjing 211169, China
    2.College of Intelligent Science Systems and Engineering, Harbin Engineering University, Harbin 150001, China
  • Online:2020-12-15 Published:2020-12-15

摘要:

为了解决移动端视觉感知模块的内存资源和硬件条件不适应卷积神经网络的快速应用等问题。以经典的DeepLab V3模型为基础,在保证分割精度的前提下,以优化卷积神经网络模型的运行速度为目标,提出一种新的轻量化网络结构。所设计的网络结构将原有非线性激活函数替换成新的Swish激活函数进行精度补偿,采用改进后的轻量化MobileNet V2结构替代DeepLab V3原有的特征提取器。实验结果表明,改进的DeepLab V3网络模型和目前精度最高的DeepLab V3+算法相比,其在维持一定精度的前提下,参数量和计算复杂度大大减小,运行速度明显提升,模型内存占用率下降了近96%,综合性能更强,更适合对分割性能要求较高的快速分割网络。

关键词: 图像语义分割, 移动端, DeepLab V3模型, 轻量化, 卷积神经网络

Abstract:

Based on the classic DeepLab V3 model, a new lightweight network structure is proposed, which aims at optimizing the operation speed of convolutional neural network model. The original nonlinear activation function is replaced with a new swish activation function for precision compensation, and the improved lightweight MobileNet V2 structure is developed to substitute for the original feature extractor of DeepLab V3 in the designed network structure. Experimental results show that the improved Deeplab V3 network model has higher accuracy, and parameters amount and calculation complexity are greatly reduced and the running speed is significantly improved comparing with the current Deeplab V3+ algorithm. The model’s memory occupancy rate is reduced by 96% and its stronger comprehensive ability makes it more suitable for fast network segmentation with high segmentation performance requirements.

Key words: image semantic segmentation, mobile terminal, DeepLab V3 model, lightweight, convolutional neural network