K Simonyan, Very Deep Convolutional Networks for Large-Scale Image Recognition, 2014
與同年比賽的GoogLeNet一樣覺得要增加辨識率就要Go Deeper
How to go deeper?
- 導入最小單位的3x3Filter
- 2個3x3Filter等價於一個5x5Filter
- 好處?
- 多了兩層ReLU提供更高的識別區別力
- 產生較少的運算子
- 2個3x3 C channel = 2 x(32C2) = 18C2
- 1個5x5 C channel = 1 x(52C2) = 25C2
What is VGG?
- 增加採樣率提升辨識結果
Contribution
- 越深越好
- 證明local response normalization沒有用
- LRN為跨feature map 進行normalization(pixel-wise),讓資料的差異性更大
reference:https://arxiv.org/pdf/1409.1556/
