An Improved Convolutional Neural Networks Framework with Flexible Channel Shift

With the popularity of mobile device terminals, it will become more and more common for visual models to run on mobile device terminals. Particularly, the accuracy and speed requirements for recognition will become higher and higher. A good design balances the accuracy and speed of models by reducing the amount of parameters and calculations so that these models are deployed easier on mobile terminals. Fortunately, FE-Net, as one of the state-of-the-art model architectures, introduces features into computation progressively through a limited feature space. However, the feature mapping is not diverse enough, which affects the applications of FE-Net on mobile terminals. To address this challenge, we propose an improved convolutional neural network framework with flexible channel shift, Fe-CsNet. It introduces a channel-shift operation to adjust the channel-wise dependencies dynamically in the FE-Net. The flexibility allows the network to better capture and exploit the feature maps. The experimental results show that our Fe-CsNet improves channel information communication via the channel shift operation, and improves performance without increasing inference speed and computational cost.