CNN is a representative algorithm of deep learning, which is a class of feedforward neural networks that includes convolutional computation and has a deep structure. A typical CNN is composed of 3 parts：
- Convolutional layers
Responsible for extracting local features in the image
This process could be understood as using a filter (the ll) to filter individual small regions of the image, and get the feature values of these small regions.
- Pooling layers
Greatly reduce the dimension of images to avoid over-fitting and reduce time complexity
In the above image, we can see that the original image is 20×20, and we decimate it with a sampling window of 10×10, eventually downsampling it into a feature map of 2×2 size.
- Fully connected layers The data processed by the convolutional and pooling layers are fed to the fully connected layer to get the final desired result
CNN is the best at image processing, it can:
- effectively downscale data volume
- effectively retain image features, in line with the principle of image processing (similar to human vision perception)
