Xception: Deep Learning with Depthwise Separable Convolutions

By Franc Chollet et al

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
1.1. The Inception hypothesis
1.2. The continuum between convolutions and separable convolutions
2. Prior work
3. The Xception architecture
4. Experimental evaluation
4.1. The JFT dataset
4.2. Optimization configuration
4.3. Regularization configuration
4.4. Training infrastructure
4.5. Comparison with Inception V3
4.5.1 Classification performance

Summary

The Xception architecture presents an interpretation of Inception modules in convolutional neural networks, proposing a novel network architecture inspired by Inception but with depthwise separable convolutions. It outperforms Inception V3 on large-scale image classification tasks and introduces a linear stack of depthwise separable convolution layers. The experimental evaluation compares Xception to Inception V3 on ImageNet and JFT datasets, showing improvements in performance, particularly on the JFT dataset. The Xception architecture is designed for efficient feature extraction and easy modification, providing a competitive alternative to existing models.