Theory 1: Deep Learning - Detailed Explanation of the LetNet Model

​
​

Event address: CSDN 21-day Learning Challenge

Introduction

LeNet is the first real convolutional neural network proposed by Y. Lecun et al. in 1998. Now LeNet now mainly refers to LeNet5, which is mainly characterized by convolutional layers and lowerThe sampling layers are combined as the basic structure of the network, which includes 3 convolutional layers and 2 downsampling layers and 2 fully connected layers.The original purpose of designing LeNet is to recognize handwritten characters and printed characters, and the effect is very good.

Network Structure

Structure Explanation:

• The C1 layer is a convolutional layer with 6 channels, which is obtained by convolution of the input image through 6 5x5 convolution checks.
• The S1 layer is a downsampling layer with 6 channels. It is obtained by the average pooling of the feature map of the C1 layer through a 2x2 window with a step size of 2, and is transformed by the sigmoid activation function.
• C3 is a convolutional layer with 16 channels, which is obtained by convolving S2 with 16 5x5 convolution kernels.
• S4 is a downsampling layer with 16 channels, which is obtained by the average pooling of the feature map of the C3 layer through a 2x2 window with a step size of 2, and using the sigmoid activation function to transform.
• C5 is a convolutional layer containing 120 feature maps, which is obtained by convolving S2 with 120 5x5 convolution kernels.
• F6 is a fully connected layer with 84 neurons, using a hyperbolic tangent activation function.
• output is the output layer with 10 neurons.

Network Process:

• k_size: represents the size of the convolution kernel
• k_num: represents the number of convolutions
• s: represents the step size

Calculation formula introduction:

• Convolution calculation formula:
  (n + 2p - k_size) / s + 1
  Where: n is the input image size
  p is the padding size;is 0;

• The special part of the above process is the process from AvgPool4 to Conv5. The input feature map size is the same as the convolution kernel size, and one-dimensional data is obtained.

Summary:

As the earliest convolutional neural network, LetNet is characterized by a small number of layers, few parameters, and fast training; although the model is relatively small, it has what a real neural network should have (although the sparrow is small and complete), in handwritingWord recognition has better results.