CNN is commonly used in the above image recognition
CNN overall architecture
Input Data, after convolution (Convolution), pooled (Pooling), convolution, pooling ... (Flattening (Flatten)) of the whole company (Fully Conneted), in which the number of layers of convolution layer, layer of the pool by their own decision
CNN—Convolution
Comprehension Filter, Stride, then the convolution process, no color for a picture, that is a height, depth Filter 1 is directly obtained by multiplying the corresponding Output, for color images, the height is 3, then the depth of Filter is 3, corresponds to the multiplication corresponding to each would like to add.
Convolution of comparison with Fully Connected
Convolution is used only nine parameters, Fully Connected use all parameters
CNN—Max Pooling
In addition there Max Pooling Average Pooling, Max Pooling maximum value is set within a well region, Average Pooling is averaged
When a sufficient number of layers obtained after the convolution of a small but deep image, and then flattened, followed by a whole company, and then reduce the error back propagation. 10 categories of final output
achieved by Keras
can actually make pictures by keeping the original size Padding, Padding = 1, is in the picture plus around 0, the size of a picture after convolution with Padding, Stride, Filter size matter