News classification
Contact us
- Add: No. 9, North Fourth Ring Road, Haidian District, Beijing. It mainly includes face recognition, living detection, ID card recognition, bank card recognition, business card recognition, license plate recognition, OCR recognition, and intelligent recognition technology.
- Tel: 13146317170 廖经理
- Fax:
- Email: 398017534@qq.com
Convolutional neural network in deep learning
Convolutional neural network in deep learning
Convolutional neural network architecture involves several concepts: convolution, activation function, pooling, and local parameter sharing.
Write a picture here
Basic architecture
The
A CNN network is usually a convolution, pooling, convolution, pooling, ..., full convergence.
convolution
As can be seen from the above figure, a 6×66×6 original image is convolved with a 3×33×3 kernel (dark red part in the figure) to obtain a 4×44×4 feature map. (At this point, stride=1, padding= 0)
Here are some concepts introduced
Stride, commonly known as the step size. Indicates the number of cells the convolution kernel has moved on the picture. When stride=1, it means one grid at a time; when stride=2, it moves two grids at a time. There are the following formulas,
OutputSize=PictureSize−KernelSizeStride+1
OutputSize=PictureSize−KernelSizeStride+1
Padding, that is, fill in 0 to expand the picture. at this time,
OutputSize=PictureSize+2×Padding−KernelSizeStride+1
OutputSize = PictureSize + 2 × Padding-KernelSizeStride + 1
Pooling
updating ...
Full connection
Updating...
Fundamental
1. Convolution, convolution is the bitwise multiplication of the original image with a well-designed matrix (commonly referred to as a filter), which results in a new matrix. For example, suppose we want to identify the rat's tail. We can design a filter similar to the one below.
Write a picture here
Next, the beginning terminates the convolution - the original image is multiplied by the filter in bits. When you recognize the mouse tail, you get a large convolution value.
Write a picture here
Conversely, when the rat's tail is not recognized, the convolution result will be very small, and it will be 0.
Write a picture here
2. activation function, also commonly referred to as the activation function kernel function may be mapped to the nonlinear type separable linearly separable. For example, in the figure below, you can use abs (absolute value function) as an activation function to separate blue and red points linearly. However, classification problems commonly use softmax as an activation function.
Write a picture here
3. Local parameter sharing, full convergence is an extremely costly operation, and local parameter sharing and pooling techniques are used in convolutional neural networks to improve this process. Local parameter sharing can be seen from the figure below. With deeper layers, g3 is affected by x1~x5.
Write a picture here
4. Pooling, another important technology is pooling, which can to some extent identify images with varying degrees of drift. For example, a picture with the face in the middle, and a picture with the face on the left (or right) of the picture, are almost indistinguishable from a convolutional neural network.
Write a picture here
Basic architecture
The
A CNN network is usually a convolution, pooling, convolution, pooling, ..., full convergence.
convolution
As can be seen from the above figure, a 6×66×6 original image is convolved with a 3×33×3 kernel (dark red part in the figure) to obtain a 4×44×4 feature map. (At this point, stride=1, padding= 0)
Here are some concepts introduced
Stride, commonly known as the step size. Indicates the number of cells the convolution kernel has moved on the picture. When stride=1, it means one grid at a time; when stride=2, it moves two grids at a time. There are the following formulas,
OutputSize=PictureSize−KernelSizeStride+1
OutputSize=PictureSize−KernelSizeStride+1
Padding, that is, fill in 0 to expand the picture. at this time,
OutputSize=PictureSize+2×Padding−KernelSizeStride+1
OutputSize = PictureSize + 2 × Padding-KernelSizeStride + 1
Pooling
updating ...
Full connection
Updating...
Fundamental
1. Convolution, convolution is the bitwise multiplication of the original image with a well-designed matrix (commonly referred to as a filter), which results in a new matrix. For example, suppose we want to identify the rat's tail. We can design a filter similar to the one below.
Write a picture here
Next, the beginning terminates the convolution - the original image is multiplied by the filter in bits. When you recognize the mouse tail, you get a large convolution value.
Write a picture here
Conversely, when the rat's tail is not recognized, the convolution result will be very small, and it will be 0.
Write a picture here
2. activation function, also commonly referred to as the activation function kernel function may be mapped to the nonlinear type separable linearly separable. For example, in the figure below, you can use abs (absolute value function) as an activation function to separate blue and red points linearly. However, classification problems commonly use softmax as an activation function.
Write a picture here
3. Local parameter sharing, full convergence is an extremely costly operation, and local parameter sharing and pooling techniques are used in convolutional neural networks to improve this process. Local parameter sharing can be seen from the figure below. With deeper layers, g3 is affected by x1~x5.
Write a picture here
4. Pooling, another important technology is pooling, which can to some extent identify images with varying degrees of drift. For example, a picture with the face in the middle, and a picture with the face on the left (or right) of the picture, are almost indistinguishable from a convolutional neural network.