ImageNet: VGGNet, ResNet, Inception, and Xception with Keras

Link: https://www.pyimagesearch.com/2017/03/20/imagenet-vggnet-resnet-inception-xception-keras/

Key concepts:

• CNN: Convolutional Neural Network
• ImageNet dataset
• ImageNet Large Scale Visual Recognition Challenge: train a model that can correctly classify an input image into 1,000 separate object categories. ImageNet challenge is the de facto benchmark for computer vision classification algorithms
• Weight serialization
• Weights of a neural network
• Pooling
• Softmax

Notes:

• Keras comes integrated with VGG16, VGG19, ResNet50, Inception V3, and Xception neural network models (look inside applications submodules)
• ImageNet: manually labeled 22 000 object categories
• ImageNet Large Scale Visual Recognition Challenge: train a model that can correctly classify an input image into 1,000 separate object categories.
• ImageNet challenge is the de facto benchmark for computer vision classification algorithms

VGG16 VGG19

• arXiv:
  • https://arxiv.org/abs/1409.1556
• The VGG network architecture was introduced by Simonyan and Zisserman in their 2014 paper, Very Deep Convolutional Networks for Large Scale Image Recognition.
• 3×3 convolutional layers stacked on top of each other in increasing depth
• The “16” and “19” stand for the number of weight layers in the network
• The smaller networks converged and were then used as initializations for the larger, deeper networks — this process is called pre-training.
• Slow to train
• Huge compared to other networks, in terms of bandwidth and disk usage

ResNet

• arXiv:
  • https://arxiv.org/abs/1512.03385
  • https://arxiv.org/abs/1603.05027
• First introduced by He et al. in their 2015 paper, Deep Residual Learning for Image Recognition
• Micro-architecture modules -> building blocks to construct the network
• ResNet50 -> 50 weight layers, based on 2015 paper

Inception V3

• arXiv:
  • https://arxiv.org/abs/1409.4842
  • https://arxiv.org/abs/1512.00567
• First introduced by Szegedy et al. in their 2014 paper, Going Deeper with Convolutions
• The goal of the inception module is to act as a “multi-level feature extractor” by computing 1×1, 3×3, and 5×5 convolutions within the same module of the network
• GoogLeNet -> Inception vN, N = version number put by google
• Keras has inception from paper of 2015

Xception

• arXiv:
  • https://arxiv.org/abs/1610.02357
• The original publication, Xception: Deep Learning with Depthwise Separable Convolutions, by François Chollet
• Smallest weight serialization

SqueezeNet

• arXiv:
  • https://arxiv.org/abs/1602.07360
• Can achieve levels of accuracy similar to AlexNet.

Classifying images with VGGNet, ResNet, Inception, and Xception with Python and Keras

CNN Image pixel sizes

• ImageNet : 224×224, 227×227, 256×256, and 299×299; however, you may see other dimensions as well.
• VGG16 : 224×224
• VGG19 : 224×224
• ResNet : 224×224
• Inception V3: 299x299
• Xception: 299×299 pixel inputs

The weights for the chosen neural network will be downloaded in the first use and will be available for later uses with Keras

Links:

• ImageNet dataset at http://image-net.org/
• ImageNet Large Scale Visual Recognition Challenge at http://www.image-net.org/challenges/LSVRC/