table of Contents
2 Convolutional Neural Network
2.2.4 Non-linear transformation layer (activation function)
1 Functional API
2 Convolutional Neural Network
This article will focus on using Convolutional Neural Networks (CNN) in keras to process images.
CNN is indeed inspired by the biology of the visual cortex .
Simply put: the visual cortex has a small number of cells that are sensitive to certain parts of the visual area
sense.
For example: some neurons are only excited on vertical edges, and some are excited on horizontal or diagonal edges .
2.1 CNN basics
2.1.1 Work flow
An overview of CNN's work means you pick an image and let it go through a series of:
-
Convolutional layer,
- Nonlinear layer
- Pooling (downsampling) layer
- Fully connected layer,
Finally get the output.
As mentioned before, the output can be the probability of a single category or a group of categories that best describes the content of the image
2.1.2 What is convolution?
Convolution refers to applying a convolution kernel to all points of a certain tensor , and generating a filtered tensor by sliding the convolution kernel on the input tensor .
Convolution completes the
extraction
of image features or information matching
. When an image containing certain features passes through a convolution kernel, some convolution kernels are stimulated.
Live, output a specific signal.
2.2 CNN architecture
If none of these layers , the model is difficult and complex pattern matching, because the network will have too much information to fill , that is, those other layers function is to highlight important information, reduce noise .
2.2.3 Convolutional layer
- ksize Convolution kernel size
- strides The span of convolution kernel movement
- padding edge padding
2.2.4 Non-linear transformation layer (activation function)
- Relu
- sigmiod
- fishy
2.2.5 Pooling layer
2.2.6 Fully connected layer
Connecting the final output with all the features, we have to use all the features to make a decision for the final classification. Finally cooperate with softmax for classification
Overall structure
3 Actual combat
Take satellite network as an example
3.1 Data classification:
import pathlib # Much easier to use than Os
3.2 Build a model
3.3 Result analysis
