基于注意力机制的模糊数字识别研究

RESEARCH ON FUZZY DIGIT RECOGNITION BASED ON ATTENTION MECHANISM

提出一种能有效提高模糊数字识别准确率的SCDM(Space-Channel Domain Mixed)模块。在进行训练时应用该模块,使中间特征图沿着空间与通道两个不同的维度生成注意力特征图,这样能够强化有效信息、削弱无效信息。在传统的卷积神经网络中通过引入该附加的模块,能有效地突出空间和通道上的特征,从传统网络中获得更高的准确率。在SVHN和MNIST两个数据集上设计实验,证明了卷积神经网络中使用该模块能够有效提高模糊数字识别的性能。

This paper proposes a SCDM module that can effectively improve the accuracy of fuzzy digit recognition. During training, using the attention mechanism, the intermediate feature map generated attention feature map along two different dimensions of space and channel, which could strengthen the effective information and weaken the invalid information. By introducing this additional module in the traditional convolutional neural network, the features on the space and channels could be effectively highlighted, and a higher accuracy rate could be obtained from the traditional network. The experiment was designed on two data sets of SVHN and MNIST, which proved that the use of this module in convolutional neural network can effectively improve the performance of fuzzy digit recognition.