1. Brief introduction
Xception is a convolutional neural network with a depth of 71 layers, relying only on depth-separable convolutional layers.
The paper explains the Inception module in convolutional neural networks as an intermediate step between regular convolution and depthwise separable convolution operations (depthwise convolution followed by pointwise convolution). From this perspective, depthwise separable convolutions can be understood as Inception modules with the largest number of towers.
The figure below is the original depth separable convolution
Improved Depthwise Separable Convolutions in Xception
The original depthwise separable convolution is a depthwise convolution followed by a pointwise convolution. The modified depthwise separable convolution is a pointwise convolution followed by a depthwise convolution .
As a new deep convolutional neural network architecture inspired by Inception, where the Inception module has been replaced by a depthwise separable convolution. Shows that this architecture, called Xception, performs slightly better than Inception V3 (for which Inception V3 was designed) on the ImageNet dataset, on a much larger image classification dataset containing 350 million images and 17,000 categories , its performance is significantly better than Inception V3. Since the Xception architecture has the same number of parameters as Inception V3, the performance is improved