摘要
为提高细粒度图像分类的精确度,提出一种基于双线性网络(Bilinear CNN)的改进方法。首先,选取结构紧密的DenseNet121卷积部分作为特征提取模块,运用改进的Relu-and-Softplus激活函数;接着,结合注意力机制引入空间注意力模块和通道注意力模块,在整体性和局部性上有效提取细节特征;并增加一层卷积层实现调整特征图维度的过渡作用,通过特征图分组策略有效降低特征向量维度减少参数;在双线性池化后采用全局最大池化层处理N个双线性特征向量,融合得到用于Softmax分类的最终向量。经实验证明,新模型的分类精确度可达到96.869%,参数量也大幅度降低,工作效率显著提高。
In order to improve the accuracy of fine-grained image classification,an improved method based on Bilinear CNN is proposed.First,DenseNet121 convolution part with tight structure is selected as the feature extraction module,and the improved Relu-and-SoftPlus activation function is used.Then,combining with the attention mechanism,the spatial attention module and the channel attention module are introduced to extract the detailed features effectively on the integral and local aspects.A convolutional layer is added to adjust the dimensionality of the feature map,and the feature map grouping strategy is adopted to effectively reduce the dimensionality of the feature vector and reduce the parameters.After bilinear pooling,N bilinear eigenvectors are processed by global maximum pooling layer,and the final vector for Softmax classification is obtained by fusion.The experimental results show that the classification accuracy of the new model can reach 96.869%,the number of parameters is also greatly reduced,and the working efficiency is significantly improved.
作者
田佳鹭
邓立国
TIAN Jialu;DENG Liguo(School of Mathematics and Systems Science,Shenyang Normal University,Shenyang 110034)
出处
《计算机与数字工程》
2021年第5期977-981,1017,共6页
Computer & Digital Engineering
基金
辽宁省教育科学规划课题“教育信息化云生态环境的架构大数据研究”(编号:JG16DB395)
辽宁省教育厅高校科研项目“基于区块链智能合约健康医疗大数据价值转移和数据共享研究”(编号:LJC202008)
国家社会科学基金艺术学重大项目(编号:18ZD23)资助。
