Preface
This article focus on the deep learning in computer vision areas organial by me. If you want to copy something, please referring this. Thank you.
1 History
The main challenges of deep learning are to solve the tasks which can be solved easily by human beings but hardly for computers.
At begining, features be got by traditional way, just like getting the color and texture of the object picture to make the computer can get the ROI. It’s called engineered hand feature by model pattern.
By the ability of the computation growing up and the GPU get more powerful use, there is a possibility to make sample units to fix the most complex model with ANN.
2 Networks
2.1 Basic unit
So let’s talk about the most excited part of the most basic function of the deep learning as the basic unit of the net.
y=f(wx+b)
x is the input of the layer, y is output of the layer. w is called the weight which can be trained by the obtained result. the wx+b is easily regarded as the liner relation, and the result of it is unlimited.If we want make it limited, we can use the activation function. The f(.) is a nonlinear function named activation function which commonly uses sigmoid function and ReLU function.
2.2 Autoencoder（AE）
Autoencoder (AE) is a kind of basic networks
2.3 Stacked Autoencoder (SAE)
2.4 Deep Belief Network (DBN)
2.5 Convolutional Neural Network(CNN)
Convolution is a kind of method to find the relation of the pixels from the input images.using a kernels can get the features from the image.