A Novel Attention Model of Deep Learning in Image Classification

Qiang Hua, Liyou Chen, Pan Li, Shipeng Zhao, Yan Li

Abstract: As the neural network becomes more and more complex, a large number of parameters are to be adjusted and more unrelated information will be generated, which is more time-consuming in model training and affects the model performance. The attention mechanic is often used to address this problem in the literature. However, the current attention models only consider the correlation within the pixel domain and channel domain individually, and the computational complexity is comparatively high. This paper proposes a new type of attention module called Pixel-wise And Channel-wise attention (PAC attention module for short) in which the hybrid correlation between the pixel and channel domain is also considered. Without increasing the number of parameters, this module can be used to calculate the correlation of the convolution feature map at any position in any layer of the convolutional neural network to achieve end-to-end training. In the experiments, the proposed PAC attention module is used in typical network structures in deep learning and has been compared to current models on some benchmark big datasets.