GoogleNet neural network structure
A module inception four parallel lines as follows: 1. The convolution of a 1 x 1, a small receptive fields convoluted convolution extracting features 2. 1 x 1 plus a convolution of a 3 x 3, 1 convolution x 1 channel inputs reduction features reduce the amount of calculation parameters, and then a 3 x 3 convolution to do a larger receptive field of a convolution 3. adding a 1 x 1 to 5 x 5 convolution function and the second 4 as a 3 x 3 convolution plus the maximum pool of 1 x 1, the arrangement characterized in the largest pool of input is changed, the 1 x 1 convolution feature extraction, and finally the four parallel lines obtained are spliced together on the channel characteristics this dimension.
InceptionNet (GoogleNet) is formed of a stack of modules Inception. InceptionNet same sized portions of the input normalized to a Block (different from vgg_block), the entire network looks more layered.
