CNN return convolution neural network directory
Previous: depth articles - CNN convolutional neural network (a) elaborate cnn convolution neural network
Next: depth articles - CNN convolutional neural network (c) the ROI Pooling and ROI Align interpolation on
In this section, elaborate pools and Anti pooling, the next section elaborate on ROI and interpolation
II. Pooling layer (pooling layer)
1. Pooling
(1) Objective of the cell layer on the input image is subsampled to reduce the computational load, memory utilization, and the number of parameters (thus reducing the risk of over-fitting). Reducing the size of the input image so that the neural network can also tolerate a certain image shift (position invariance). Pooling sampling dimensionality reduction is an efficient sampling.
(2). Convolution layer neurons pooled layer similar element, is connected to the surface layer of the output characteristic, only pool layer is locally connected by pooling core (Kernel). Likewise, the definition must be pooled core (Kernel) size, pace and fill type. However, there is no pooling neuron weights, all of it to do is to use aggregation functions (such as max or mean) output polymerization
(3) of the cell layer can be extracted feature dimensionality reduction information, one feature map (feature map) becomes smaller, simplified network computational complexity, and to avoid over-fitting occurs to some extent; the one hand compression feature, extract the main features.
(4) The pool unit shift invariance, even a small displacement of the image, the extracted features still remains unchanged. Because of the enhanced robustness to displacement, so that the target can ignore the inclination, relative position change of the rotation or the like, in order to improve the accuracy.
(5) Method of cell:
①. The maximum pooling (max pooling)
Save the maximum value of the nuclear pool of a sliding window, in order to achieve the effect of dimension reduction
In the largest pool of two-dimensional functions:
tf.nn.max_pool(value, ksize, strides, padding, name=None)
value: input data, typically in the back contact layer pooled convolution layer, the input is typically feature map, still (b, h, w, c) such a shape
ksize: the pool size of the core window, taking a four-dimensional vector, generally: [1, height, width, 1], because the do not want pooled in batch and channels, so that 1 is set to two dimensions.
strides: the pool of the nuclear stride, pace convolution kernel and the like, are generally [1, h, w, 1]
padding: filled, the convolution kernel and the like, can be taken "SAME" or "VALID". In pooling in, padding usually selected "VALID", because pooling itself is to take the loss of data in order to achieve the effect of dimensionality reduction. On this basis, it lost more edge data is not very important, but also is within an acceptable range.
②. The average pooled (average pooling)
After storage of the average cell nucleus of a sliding window sum value is rounded down.
In the two-dimensional average pooled function:
tf.nn.avg_pool(value, ksize, strides, padding, name=None)
Same () parameters meaning max_pool.
2. Han'ikeka
(1). Pooling is pooled anti-reverse operation is unable to restore the original data by all the results pooled. Because the pool of the process is to retain only the main message, discarding some of the information.
(2) If you want to recover all the information these key information from pooling, there is information missing, then it can only be achieved through complete the greatest degree of information to fill the seats. Of course, if the Laplacian downsampling ideas or information can be restored from the pool of all the information (in the Laplace downsampling, apart FIG enter and output, there is a Laplace FIG Si matrix, for use when the sample and re-stored).
(3) There are two types of cell: cell and the maximum average pooling, which also require anti-pooling corresponding thereto.
①. The average tank and anti-tank average of
First restore the original size, and then each value results are pooled fill the raw data area corresponding to respective positions.
②. Max Pool and de-largest pool of
Records coordinate position of the maximum value of the activation process of the pool, and then restore the original size, when the anti-pooling, the pooling process as long as the maximum value of the location coordinate activation of the activation, other values set to zero. Of course, this process is only an approximation. Because the process of the pool, except the position of the maximum value, other values are not all zero.
CNN return convolution neural network directory
Previous: depth articles - CNN convolutional neural network (a) elaborate cnn convolution neural network
Next: depth articles - CNN convolutional neural network (c) the ROI Pooling and ROI Align interpolation on
