Inception DenseNet With Hybrid Activations For Image Classification
Jinming Zhang, Zuren Feng · 2019
The familly of Inception network and DenseNets are some of the most successfull CNN architectures proposed in recent years. Researchers of Inception-ResNet combined the residual connections and the Inception architecture by replaceing the tensor concatenation step of the Inception modules using residual connections. Later, the DenseNet archicture demonstrates that using dense connections (implemented via filter concatenation) instead of residual connections can also be very effective to train deeper networks. This gives the problem of are there any performance on combining the Inception-like blocks with dense connection. In this paper, we design a network architecture by embedding the Inception-like blocks into DenseNet architecture, which is called Inception-DenseNet architecture. Another innovation is that our inception-like blocks introduce hybrid activation operations, which is different from previous inception blocks. Visualization experiments show that this hybrid activation mode can make more flexible response to object semantic regions. The performance of our net models are compared with the current best net models such as DenseNets, ResNets on several image datasets. The experiment results indicate that our Inception-DenseNet can obtain the same or better classification accuracy while use a smaller number of trainable parameters.