Journal of Computer Applications ›› 2020, Vol. 40 ›› Issue (3): 645-650.DOI: 10.11772/j.issn.1001-9081.2019081425

• Artificial intelligence • Previous Articles     Next Articles

Image classification algorithm based on lightweight group-wise attention module

ZHANG Panpan1,2, LI Qishen2, YANG Cihui2   

  1. 1. School of Information Engineering, Nanchang Hangkong University, Nanchang Jiangxi 330063, China;
    2. Key Laboratory of Jiangxi Province for Image Processing and Pattern Recognition(Nanchang Hangkong University), Nanchang Jiangxi 330063, China
  • Received:2019-08-16 Revised:2019-10-15 Online:2020-03-10 Published:2019-11-06
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61662048), the Graduate Special Innovational Foundation of Nanchang Hangkong University (YC2018030).

基于轻量级分组注意力模块的图像分类算法

张盼盼1,2, 李其申2, 杨词慧2   

  1. 1. 南昌航空大学 信息工程学院, 南昌 330063;
    2. 江西省图像处理与模式识别重点实验室(南昌航空大学), 南昌 330063
  • 通讯作者: 李其申
  • 作者简介:张盼盼(1993-),女,江苏宿迁人,硕士研究生,主要研究方向:图像处理、模式识别、深度学习;李其申(1975-),男,河北衡水人,副教授,博士,主要研究方向:图像处理、模式识别;杨词慧(1979-),男,江西九江人,副教授,博士,主要研究方向:图像处理、模式识别。
  • 基金资助:
    国家自然科学基金资助项目(61662048);南昌航空大学研究生专项创新资金资助项目(YC2018030)。

Abstract: Aiming at the problem that the existing neural network models have insufficient ability to characterize the features of classification objects in image classification tasks and cannot achieve high recognition accuracy, an image classification algorithm based on Lightweight Group-wise Attention Module (LGAM) was proposed. The proposed module reconstructed the feature maps from the channel and space of the input feature maps. Firstly, the input feature maps were grouped along the channel direction, and channel attention weight corresponding to each group was generated. At the same time, ladder type structure was used to solve the problem that the information between the groups was not circulated. Secondly, the global spatial attention weight was generated based on the new feature maps concatenated by each group, and the reconstructed feature maps were obtained by weighting the two attention weights. Finally, the reconstructed feature maps were merged with the input feature maps to generate the enhanced feature maps. Experiments were performed on the Cifar10 and Cifar100 datasets and part of the ImageNet2012 dataset with using the classification Top-1 error rate as the evaluation indicator to compare the ResNet, Wide-ResNet and ResNeXt enhanced by LGAM. Experimental results show that the Top-1 error rates of the neural network models enhanced by LGAM are 1 to 2 percentage points lower than those of the models before enhancing. LGAM can improve the feature characterization ability of existing neural network models, thus improving the recognition accuracy of image classification.

Key words: attention mechanism, image classification, channel attention, spatial attention, group convolution

摘要: 针对图像分类任务中现有神经网络模型对分类对象特征表征能力不足,导致识别精度不高的问题,提出一种基于轻量级分组注意力模块(LGAM)的图像分类算法。该模块从输入特征图的通道和空间两个方向出发重构特征图:首先,将输入特征图沿通道方向进行分组并生成每个分组对应的通道注意力权重,同时采用阶梯型结构解决分组间信息不流通的问题;然后,基于各分组串联成的新特征图生成全局空间注意力权重,通过两种注意力权重加权得到重构特征图;最后,将重构特征图与输入特征图融合得到增强的特征图。以分类Top-1错误率作为评估指标,基于Cifar10和Cifar100数据集以及部分ImageNet2012数据集,对经LGAM增强之后的ResNet、Wide-ResNet、ResNeXt进行对比实验。实验结果表明,经LGAM增强之后的神经网络模型其Top-1错误率均低于增强之前1至2个百分点。因此LGAM能够提升现有神经网络模型的特征表征能力,从而提高图像分类的识别精度。

关键词: 注意力机制, 图像分类, 通道注意力, 空间注意力, 分组卷积

CLC Number: