Deep neural network
The deep neural network is
a multi-layer deep learning model that simulates the construction of human
brain neurons. A deep neural network is a neural network with a certain level
of complexity, a neural network with more than two levels of complexity. Deep
neural networks use sophisticated mathematical modeling to process data in
complex ways in the field of hyperspectral image classification, the deep
neural network model represented by CNN has been widely used[45]. In particular,
input passage through different layers of pattern recognition and simulated
neural connections. This section takes CNN as the representative of deep neural
networks and introduces its calculation principle and action mechanism. The
term "deep learning" is also used to describe these deep neural
networks, as deep learning represents a specific form of machine learning where
technologies using aspects of artificial intelligence are the way to classify
and order information outside of simple input and output protocols.
CNN was first proposed by
computer science researcher Yann LeCun in 1980. The Convolution Neural Network
(CNN) is like all other general neural networks. Convolutional networks are
specialized types of neural networks that use convolution in place of general
matrix multiplication in at least one of their layers[46]. Convolutional
Nuclear Network (CNN) is one of the most widely used deep learning-based
methods for visual data processing. The use of CNN for HSI classification is
also visible in recent work. Convolution is the process of extracting
features[47]. The inputs are the image matrix and the kernel matrix or filter,
which will be used to extract the feature[48]. A feature map will be available
as output. One by one filters can be used to extract one type of feature. Such
as edge detection, image and video recognition, recommender systems. CNN
typically includes the convolutional layer, the pooling layer, and the
fully connected layer[2][49] .
Figure 2.2 Operation mode of convolutional layer
The
convolution layer uses filters that perform convolution operations as it is
scanning the input concerning its dimensions. Its hyperparameters include the
filter size and stride[49]. The resulting output is called a feature map or
activation map. The pooling layer is a down sampling operation. typically
applied after a convolution layer which does some spatial invariance. In
particular, max and average pooling are special kinds of pooling where the
maximum and average values are taken respectively. The filter in the
convolutional layer will perform convolution calculations on pixels through a
sliding window. In addition to the
size convolution window, the size of the
output feature is different from the original feature as shown in figure 2.2.
Use the convolution window with the size of
to convolve for a
pixel input image and the output image is
pixels. When the step size is
, the convolution
window slides
times on the original image. The resulting
output feature map, also known as a factor map or activation map, has all the
elements calculated using the input layer and filtering. Different kernel
functions can be defined in the convolution kernel, and the kernel function and
the calculation result of the pixel value matrix of the corresponding window in
the input feature graph are used as the pixel value of the output feature
graph. Which has all the elements calculated using the input layer and
filtering and generated output feature map is also called a feature map or an
activation map[50]. Different kernel functions can be defined in the
convolution kernel. The kernel function and the calculation result of the pixel
value matrix of the corresponding window in the input feature map are used as
the pixel values of the output feature map.
Figure 2.3 Operation mode of pooling layer
The
pooling layer is usually used after the convolutional layer to discard pixel
values. Pooling methods are usually divided into two maximum pooling and
average pooling[51]. As shown in Figure 2.3, the maximization pooling
operation is to select the maximum value among the pixel values in the
specified size window to retain the maximum value and discard the remaining
pixel values. The average pooling operation is to average pixel values in the
specified size window and use average to replace all pixel values in the window.
The pooling operation feature will reduce the
size of the image which can simplify the calculation network to compress the
feature and reducing the complexity of image processing. Proper pooling
operation in the CNN algorithm will improve the ability of the convolution
function to extract features, which can reduce the amount of calculation and
improve the performance of feature extraction. Neural network the fully
connected layer is a part, where the neurons of each layer of the network are
fully connected with the neurons of the connected layer [50]. As shown in Figure 2.4. A fully connected layer is
usually used at the end of the network to connect the hidden layer and output
layer to connect the hidden layer to the output layer.
Figure 2.4 Fully connected layer connection mode
The
number of neurons connected to the output layer must be equal to the number of
categories of the predicted data. When the number of full connection layers is
large, the neurons are usually randomly inseminated by Dropout to reduce the
effect of overfitted normalization.
0 মন্তব্য(গুলি):
একটি মন্তব্য পোস্ট করুন