计算机与现代化 ›› 2021, Vol. 0 ›› Issue (12): 79-84.

• 图像处理 • 上一篇    下一篇

深度连接的超轻量化子空间注意模块

  

  1. (浙江工商大学信息与电子工程学院,浙江杭州310018)
  • 出版日期:2021-12-24 发布日期:2021-12-24
  • 作者简介:张宸逍(1997—),男,浙江温州人,硕士研究生,研究方向:计算机视觉,E-mail: 19020090020@pop.zjsu.edu.cn; 潘庆(1997—),女,硕士研究生,研究方向:计算机视觉,E-mail: 1454556150@qq.com; 王效灵(1967—),男,教授级高级工程师,博士,研究方向:计算机视觉,E-mail: wangxiaoling@zjgsu.edu.cn。
  • 基金资助:
    浙江省重点研发计划项目(2018C01084)

Deep Connected Ultra-lightweight Subspace Attention Mechanism

  1. (School of Information and Electronic Engineering, Zhejiang Gongshang University, Hangzhou 310018, China)
  • Online:2021-12-24 Published:2021-12-24

摘要: 针对紧凑型卷积神经网络在部署现有注意力机制存在计算量或参数开销大的问题,提出一种改进的超轻量化子空间注意模块。首先,深度连接的子空间注意模块(Deep Connected Subspace Attention Mechanism, DCSAM)划分特征图为若干特征子空间,为每个特征子空间推导不同的注意特征图;其次,改进特征子空间进行空间校准的方式;最后,建立前后特征子空间的连接,实现前后特征子空间的信息流动。该子空间注意机制能够学习到多尺度、多频率的特征表示,更适合细粒度分类任务,且与现有视觉模型中的注意力机制是正交和互补的。实验结果表明,在ImageNet-1K和Stanford Cars数据集上,MobileNetV2在参数量和浮点运算数分别减少12%和24%的情况下,最高精度分别提高了0.48和约2个百分点。

关键词: 紧凑型, 注意力机制, 深度连接, 特征子空间

Abstract: In order to solve the problem of large computation or parameter overheads in deploying the existing attention mechanism of compact convolutional neural network, an improved ultra-lightweight subspace attention mechanism is proposed. Firstly, the deep connected subspace attention mechanism(DCSAM) is used to divide the feature map into several feature subspaces, and deduce different attention feature maps for each feature subspace. Secondly, the spatial calibration method of feature subspace is improved. Finally, the connection between the front and back feature subspaces is established to make the information flow between the front and back feature subspaces. The subspace attention mechanism enables multi-scale and multi-frequency feature representation, which is more suitable for fine-grained image classification. The method is orthogonal and complementary to the existing attention mechanisms used in visual models. The experimental results show that on ImageNet-1K and Stanford Cars datasets, the highest accuracy of MobileNetV2 is improved about 0.48 and 2 percent points when the number of parameters and floating-point operations are reduced by 12% and 24% respectively.

Key words: compact, attention mechanism, deep connection, feature subspace