Very Deep Convolutional Networks for Large-Scale Image Recognition

[chullhwan-song]

https://arxiv.org/abs/1409.1556

[chullhwan-song]

• Very Deep Convolutional Networks for Large-Scale Image Recognition (VGG network)

  특징

  • CNN
  • 224x224 input RGB image
  • very small 3×3 filters in all convolutional layer
  • case vgg16 구조
    • (2 conv layer + 1 max pooling layer)x2+(3 conv layer + 1 max pooling layer)x3 + 3개의 fc layers+softmax layer
  • relu activate function

    depth to 16–19, weight layers.

    

  • vgg 16, vgg 19
  • alexnet과 마찬가지로 fc layer에서 deep feature
    • sparse & semantic
  • alexnet보다 약 10%이상의 성능 향상을 보여줌
  • 매우 강건한 network

    성능