1968, Hubel on the study of the visual cortex cells of cats, put forward the concept of receptive filed, the visual cells can be divided into simple cells and complex cells, respectively, the range of the field of perception, on the basis of biology, the study of two-dimensional image convolution neural network.
Traditional image classification: Feature extraction + feature expression + Classification CNN sets these methods together,
One, convolutional neural network characteristics
1. Local Accepted domains
Feel the width of the wild, the retina, the M layer. M+1 layer, feeling wild outside the range is not responding,
2. Weight sharing
Share the same weights and translate them.
Two, convolutional neural network structure
Typical structure
1. Convolutional layer (convolution)
Image is a two-dimensional discrete signal, for the image, convolution is a filtering process, convolution function of the convolution weight is different, and image processing effect is different, the use of horizontal gradient convolution core results in the horizontal direction of the largest response.
The hidden layer is composed of several feature graphs, the weight of hidden layer can be expressed as the target feature graph, the Meta feature graph, the target pixel horizontal position, and the vertical position of the target pixel 4D tensor,
Convolutional cores have a BP algorithm, and each network layer has multiple convolution cores. Through convolution operation, the neural network extracts different input features, the bottom convolution kernel obtains the edge, line, angle and feature of the image, and the high-level convolution kernel gets more complicated characteristics.
After the convolution is not yet classified, the volume of the convolution is huge, such as the 128x128 pixel image, the convolution core size is 8x8. Then there are (128-8+1) ^2 convolution cores, so the next layer is the pool layer.
The pool layer accumulates feature statistics in different locations, such as the ability to calculate the eigenvalues or maximums of an area in an image, which is much lower than the original feature image.
2. Activation layer
The activation function is a nonlinear function, and the nonlinear activation function is possible for the existence of billions of cells.
3. Classification Layer
In the CNN classification, the Common classification function is multiple logistic logistic regression model, Softmax regression model. It is a classification model based on probability, which is optimized by minimizing the negative logarithm likelihood function.
DL Learning notes-CNN related knowledge