Artificial neural networks
After all, whether voice, text or image data are in a computer by a number of digits, and the neural network structure is composed of layers, interconnected by weight. Each layer has a plurality of neurons, and the multiplication value and offset data for each of the right, and the resultant value into the function calculates and outputs. Output = f (w * Input + b).
With pictures, for example, for each data are processed, you can be n (n data we want) new data. Suppose we need to image points n value classes, each image after the treatment can produce n data, we can adjust the weights and offset values for each size and suitable processing functions, we can so that the same types of pictures are output similar characteristics. For example, ten values of all fruits produced after the first treatment is a maximum, after a new picture to be processed, if a maximum data obtained first, then we can say that this picture is a fruit.
Adjust the value of each weight and bias of the process is to learn. We need a lot of time learning the training set
Artificial neural networks limitations
But we know that a picture contains a lot of pixels, each pixel contains one or more channels, through one or more layers of processing, and each one has a lot of neurons, each data has a specific weight and offset value, for a 800x800 image, the number of each of the hidden nodes are sections 300,200,100, then the number of parameters is up to hundreds of millions. Too many parameters results in the calculation slow down, but it will also cause over-fitting. To this end we have drawn convolution neural network
Convolution neural network
In convolutional neural network, there will be a convolution layer and downsampling layer, the purpose of convolution is to signal enhancement, noise reduction. Downsampling layer reduces training parameters and the degree of over-fitting.
Different convolution kernel may be different from the image processing, such as filtering, gradient processing. Grew up in step 1 will result in reduced maintenance
down-sampling process often uses a pool of commonly used methods are pooled in two ways, mean and maximum pool pooling.
The space between the hidden layer and the hidden layer decreasing resolution, in order to detect more characteristic information of each layer will increase the number of planes
The internet
-NX input layer (layer convolution (usually three) - pooled layer) - Full connection layer
cnn function commonly used in tensorflow
dimensional convolution tf.nn.conv2d (input, filter, strides, padding, use_cudnn_gpu = None, name = None)
Maximum pooled
tf.nn.max_pool (value, ksize, strides, padding, naome = none)
average pooled
tf.nn.avg_pool (value, ksize, strides, padding, naome = none)
