参考文章：Han X, Leung T, Jia Y, et al. MatchNet: Unifying feature and metric learning for patch-based matching[C]. computer vision and pattern recognition, 2015: 3279-3286.
会议水平：CVPR 2015 （本家大哥贾扬清指导韩旭峰完成的，必须一读）
方法非常的简单粗暴，完全不用看论文....
这里有一点还是值得学习的就是韩旭峰将各个通道的特征图全部给打印出来了：
Visualization for the activations in response to an example input patch at different layers in the feature network. The input
64 × 64 patch is shown at the top. For each layer, we tile its K H × W activation maps to form a 2D image. H, W and K are
the height, width and depth of the 3D activation array respectively. Red margins separates these tiles. Pseudo-colors in the tiles represent response intensity. Border artifacts may occur, but we keep our padding scheme, which retrains half of the information on the original border