how to change the AlexNet into FCNs ?

标签：

How to change the AlexNet into FCNs ?

FCNs is a network that only contain convolution layers and no fc layer at all. It‘s structure can be shown as the following figures:

This image from the paper : <Fully Convolutional Networks for Semantic Segmentation>  CVPR 2015.

　　It could locate the location of object target perfectly as shown in above images and it doesn‘t need to resize the resolution of input images, which is the mostly different from traditional CNNs. First, Let‘s review some related network parameters about AlexNet, related structure can be shown as following:

　　As we can see from the above figure, the input of images must be resized into a fixed resolution, like 224*224, due to the existance of fc_layer. The specific pipeline could be found in this blog, web link: http://blog.csdn.net/sunbaigui/article/details/39938097

　　The output of Conv 5 is: 6*6*256, we want to obtain the final results: 1*1*1000 (take the 1k classes for an example). How could we use the middle Conv 6, Conv 7, Conv 8 layers to bridge the two results ? Do we need the pool layers added ? How to set the middle parameters in each layers ? Does it really work ?

　　Let‘s do it now.

how to change the AlexNet into FCNs ?

标签：

原文地址：http://www.cnblogs.com/wangxiaocvpr/p/5421536.html