Article Directory
TF Layers Tutorial: Building a convolution neural network
About convolution neural network
Convolutional neural network (CNNs {Convolutional Neural Networks}) is the current user image classification task in the forefront of the model. CNNs raw pixel data of an image applied to a series of filters, and to extract features of a higher level of learning and use of these features of the model images are classified. CNNs mainly comprises the following three components:
-
Layer convolution, the convolution kernel which indicates the number of applications in the image. For the sub-areas of the image, the convolution layer performs a series of mathematical transformations, so that the output value of the feature maps. Use convolution layer under normal circumstances ReLU as activation function to allow the introduction of nonlinear transformation model.
-
Cell layer, which is a convolution layer extracted image data is sampled at , is to be reduced dimension feature map, thereby reducing the calculation time. Commonly used algorithm is the largest pool pool algorithm, which extracts the pixel value of the window is a pool (eg, 2x2- pixel block) of the maximum value, the pixel values of the other sub-region were discarded.
-
Dense (fully connected) layer, after sampling and the pooled layers convolution layer, fully connected layers can be classified feature. More specifically, in the fully connected layers, each layer node are connected to nodes on the layer.
Generally, CNN feature is extracted by a convolution multilayer module. Each module contains a convolution layer, followed by a reservoir layer. Followed by a final layer of a multilayer or later full convolution module to obtain connection layer classification result. Number of the last full of CNN connection layer node is equal to the number of all possible classes of prediction task, and these nodes which will produce a value between 0 and 1 (the layer all nodes values after softmax activation function sum 1). These values can be interpreted as softmax enter the picture most likely probability which category belongs.
Note: For a more in-depth understanding of CNN's architecture, see the Stanford University course materials convolution neural network
Reference article: TF Layers Tutorial: Build a convolution neural network
