- A Neural Algorithm of Artistic Style
- 作者:Leon A. Gatys, Alexander S. Ecker, MaAhias Bethge
- Neural Networks
- Convolutional Neural Networks
- VGG 19
- h(x)
- Sigmoid: 在多層網路中容易發生 gradient vanish
- Rectified Linear
- shared weight: 用同一種 weight ,降低模型複雜度
- Stride
- Padding
- Pooling Layer
- Maximum Pooling
- Average Pooling
- Receptive Fields
- Very Deep Convolutional Networks for Large-scale Image Recognition.
- 19 (+5) layers
- 16 Convolutional layers (width=3, height=3)
- 5 Max-pooling layers (width=2, height=2)
- 3 Fully-connected layers
- 人類
- VGG19
Note: BP 修正的是畫布而非 weights
- ”Style” is position-independent
- 抽取過程:隨機將部分圖片遮掉
- "G"
- p: 欲抽取 "內容" 的畫作
- a: 欲抽取 "畫風" 的畫作
- x: 欲渲染的畫布
Note:
Ltotal = alpha * Lcontent + belta * Lstyle => alpha, belta 控制畫作偏寫實或抽象