Return to Deep Learning Network model development history of classic catalog
Previous: depth articles - classic history of Deep Learning Network model development (b) elaborate structure and features AlexNet
Next: depth articles - classic History of Deep Learning Network model (d) elaborate structure and features VGG16
In this section, ZF-Net elaborate structure and characteristics, the next section elaborate structure and features VGG16
Papers Address: Visualizing and Understanding Convolutional Networks
II. Network Classic (Classic Network)
3. ZF-Net
ZF-Net is a 2013 winner of ImageNet classification tasks, its network structure is not much improvement, but adjust the parameters, the large convolution kernels removed a bit smaller, but performance than AlexNet improved a lot.
(1) Network Description:
Network architecture:
Deconvolution, anti-tank structure:
ZF-Net uses deconvolution inverse pooled to view the extracted features results in FIG. These effects and with reference to FIG perform parameter adjustment.
. (2) ZF-Net features:
①. ZF-Net is based on fine-tuning AlexNet
a. AlexNet difference in performance with the sparse two connecting structure of the GPU, and ZF-Net only a dense structure is connected to a GPU.
. AlexNet first layer B changes: the size of the filter is about to become , and the stride becomes .
②. ZF-Net compared with AlexNet, using a front layer of less convolution kernel, and smaller steps, it retains more features.
③. Using deconvolution, visualization feature map. As can be seen by the feature map, characterized in hierarchical architecture. In front of a physical layer learning profile, edge, color, texture and other characteristics, the latter layer learning features and abstract class dependent. The more high-level features that used to make the classification performance is better; and low-level features, the fight for better box performance.
④. By blocking find the absolute key parts of an image category. Experiments described depth increases, the network can learn to distinguish more features.
⑤. When network training, low-level parameters of fast convergence, more senior, need longer time to train, to converge.
Return to Deep Learning Network model development history of classic catalog
Previous: depth articles - classic history of Deep Learning Network model development (b) elaborate structure and features AlexNet
Next: depth articles - classic History of Deep Learning Network model (d) elaborate structure and features VGG16