Xception

Xception is a convolutional neural network architecture built around depthwise separable convolutions. Instead of learning spatial and channel relationships in one standard convolution, it processes spatial patterns within each channel and then combines channels with pointwise operations. This can use parameters and computation more efficiently while maintaining strong image-recognition performance. The name reflects the idea of an extreme version of the Inception design. Xception has been used for classification and as a feature extractor in other vision systems. Results depend on input resolution, training data, augmentation, initialization, and transfer-learning choices, and newer architectures may offer better speed or accuracy for a particular deployment.