Global Attention Augmentation Ghost Module: More Features from Lightweight Global Attention Extraction

Recently, in order to deploy neural networks on mobile devices, many studies have focused on reducing the number of parameters and computational complexity of neural networks. However, most existing methods do reduce the computational complexity of deep neural networks, but also greatly sacrifice their performance. To maintain the relative balance of computational complexity and performance of deep neural networks, this paper proposes a Global Attention Augmentation Ghost(GAAG) module, which decreases the number of parameters while bring performance improvements. Analyzed the network architecture of the Ghost module, we empirically show it is waste of computational resources that cheap operation in Ghost module produces more feature maps by linear transformations, only increasing the width of convolution neural network instead of extracting more useful feature information, and in a convolution layer composed of multiple feature blocks, the circulation of channel information is essential to better integrate the information of each feature blocks. Therefore, we propose a lightweight long-range dependency extraction block instead of cheap operation, which increases the ability to extract the non-local information of the model while keeping the computational cost almost invariable. Furthermore, we combine channel shuffle and channel attention to promote the fusion of local and non-local information. The proposed GAAG module is efficient yet effective and can be flexibly plugged into existing convolutional neural networks. Experiments conducted on the benchmark demonstrate that the GAAG module can perfectly replace the traditional convolutional layer in the baseline model. We extensively evaluate our GAAG module on image classification and object detection with backbones of ResNets. The experimental results show that our GAAG module can keep a good balance between lightweight and high performance compared with the similar model.