Innovation (small convolution kernel of nuclear Koike, the number of layers deeper, become fully connected convolution)
- AlexNet improvement of using smaller stride and convolution convolution kernel convolution in a first layer
- Multiscale (training and testing, the use of pictures at different scales (of course, is based on an isotropic scaling minimum edge delineation standard))
- Layers deep
- Testing phase, will connect three full turns three convolution layer to fit any size image input (where the final score map is averaging)
Different levels of network structure
As can be seen from the figure, the structure still continued AlexNet 5-3 (5 convolutional fast, three fully connected, the last of which is the output)
Experimental results
- LRN useless
- Deepen the network, it can improve the accuracy
- Multi-scale input (training) can improve accuracy
- Multi-scale test also can improve the accuracy of (multi-scale training, multi-scale shake-scale test collectively)
- Fusion dense evaluation, and multi-scale evaluation good effect
dense evaluation indicates no crop of 224 × 224, directly into the network, connected to the full convolution
Feel the size of the wild
Pixels on the convolutional neural network wherein FIG mapped back on the original size
Two issues
Q1: Why a 3 × 3 convolution kernel
(1) decrease the amount of parameters, generating the same size when the characteristic diagram, three 3 × 3 convolution corresponds to a 7 × 7 convolution of:
. 3 × (C × . 3 × C ×. 3) 27C = 2
. 7. 7 × × × C C 49C = 2
(2) active layer increases, the nonlinear increase
Q2: 1 × 1 action convolution kernel
(1) A way to increase the decision of nonlinear elements
(2) to adjust the dimensions of the network or reduced dimensions Augmented