1. Related Theories and Algorithm Models
The basic principle of machine learning is to accurately extract the pattern or ‘model’ of a particular piece of data from a large amount of data. Then use it to classify new information, known as classification. This is machine learning. The task of machine learning is not to tell the machine exactly what to do, to teach how to do it. This is done by writing algorithms. In this case of machine learning and computer vision, many great algorithms have been applied to image processing. Algorithm means the logical steps to do a task. As a special multi-channel image, hyperspectral data does not have more complex features and information than ordinary images of wave bands, but has more complex features and information in the spatial scene. As a result, a large number of related algorithms have emerged and these algorithms are known as machine learning algorithms and cannot be separated from the basic principles of computer vision algorithms. This chapter deals with the relevant feature extraction algorithms and relevant classification algorithms engaged in the research content of this paper. Based on the principle, the relevant algorithm is deduced for the feature extraction method of hyperspectral image based on visual salience and the classification method of deep forest hyperspectral image based on attention mechanism.
1.2 Related feature extraction algorithm
Working
with data sets of hundreds of features is becoming quite common these days. If
the number of observations stored on a data set becomes the same number of
features. But it is likely that the machine learning model may suffer from
overfitting[27]. To avoid these kinds of problems regulation or level reduction
strategies need to apply feature extraction. Feature extraction techniques have
also led to other kinds of advantages, such as accuracy improvements, speed up
in training dataset overfitting risk reduction, improved visualization of the
dataset and increase in the explain ability of this model. Feature extraction
targets to reduce the number of features in the data set by creating new and
flowing features from existing features and discarding original features. The
generated low-dimensional data set contains most of the information of the
sample [6]. The PCA method is one of the most used linear dimensionality
reduction technique. PCA is able to do this by maximizing variances and
minimizing the reconstruction error by the pair looked at the distance. In PCA,
the original data is projected on a set of orthogonal axes and spaced according
to the importance of each axis. This section will first introduce the concepts
related to computer vision features, and introduce the salience feature
extraction algorithm, PCA and other feature extraction algorithms used in this
study. When using a PCA, we take the original data as input and try to find a
combination of input features that can minimize the distribution of the
original data in the best way possible to reduce its actual dimensions.
1.3 Computer vision feature
extraction
Computer
vision is an interdisciplinary studies field that aiming to develop technology
to help computers "see" and "understand" the contents of
digital images such as photos and videos like human beings[28]. Feature
extraction is an interior component in computer vision pipelines. In fact, the
deep learning model explores useful features to clearly define the surrounding
objects. A feature in computer philosophy is the measurable piece of data in
your image that is unique to this particular object. It can be a different
color or a specific shape of an image, such as a line, an edge, or an image
section. A good feature is used to distinguish them from each other. In general,
a particular problem in vision can be easily solved by a hand-designed
statistical method, while other new problems may require a large and complex
set of generalized machine learning algorithms. With the development of
artificial intelligence, computer vision is at an extraordinary moment of its
development, playing a key role in such fields as medical imaging[29] ,
automobile safety[27], and biometric identification [29].
The
feature is defined as an interesting part of the image and features are used as
an early point for many computer philosophy algorithms. Machine learning,
pattern recognition, and image processing feature extract begin with a common
set of data measurements. Computer vision features are features or maps that
can be easily understood by the machine as human vision can be extracted by
computer vision algorithms, such as texture features[30], salient features[31],
and contour features[27] that we can understand. the texture is the feature
used to divide images into important regions and to classified those areas.
texture provides an image color or important spatial arrangement information.
On topographic maps, contours represent the shape of the land. Features of the
landscape that are useful to know are elevation and steepness. hills, valleys,
depression, gullies, ridges, etc. Salient visual features are the defining
elements that distinguish one target from another. They are key pieces of
distinct information that facilitate the recognition of an image, object,
environment, or person. Focusing on computer vision, the number of use-cases
for applying AI that performs at human-level or better is increasing
exponentially, given the fast-paced advances in Machine Learning. However, high
innovation potential does not come without challenges.
1.3 Saliency theory Boolean mapping
Boolean
functions are known as Boolean expressions and Boolean functions are expressed
by algebraic expressions. Which consists of binary variables, the constants 0
and 1, and the logic operation symbols. Consider the following example. An
image is features by a set of binary images, which are generated by randomly hyperspectral
image processing of thresholding the image’s color channels. Based on a gestalt
principle of figure-ground segregation, BMS computes saliency maps by analyzing
the topological structure of Boolean maps. BMS is simple to implement and
efficient to run. Gestalt psychology studies show that size, envelopment,
concavity, and other factors will affect the isolation of image background
[31]. Boolean Map-based Saliency (BMS) is based on this principle, which uses
known global topological cues that are helpful to perceive image background
isolation to detect Saliency regions. Boolean Map-based Saliency model (BMS),
which leverages global topological cues that are known to help in perceptual
figure-ground segregation BMS is also shown to be advantageous in salient
object detection. To measure the surround, BMS identified the image by the set
of bullion maps. In BMS, attention maps are efficiently counted by binary image
processing techniques to activate the pixel region with closed external
contours on a given Boolean map. Then randomly sample saliency is modeled as
the expected attention level given to the Boolean map set. Then, according to
the given randomly sampled Boolean mapping set, the significance is modeled as
the desired target, and finally, the final significance region is obtained
according to multiple Boolean subgraphs[31].
1.4 Principal component
analysis
PCA
comes from the first letter of its three English words is Principal component
analysis (PCA). First, the data is preprocessed and then the helpful error
matrix of matrix X is counted. Then, the eigenvalues and eigenvectors of the
differential matrix are calculated. Finally, a mapping matrix is established
based on the eigenvalues and eigenvectors. PCA is mainly used for face
recognition, anomaly detection, and hyperspectral image so on. To put it more
plainly dimensionality reduction. Other expanded applications have done the
corresponding work on this basis. In real data analysis work, we must also
require the complex multidimensional data to be inked, and then some machine
learning models need to be used for training it. the direction of the axis in
which the data is viewed when viewed from the horizontal or vertical axis. As
the size of the data increases, so does the difficulty of visualizing it and
performing calculations. Therefore, reducing the size of data mainly aims to
achieve two purposes: to remove redundant size and to retain only the most
important features[33]. PCA (Principal Component Analysis) is a commonly used
data analysis method. PCA transforms the original data into a set of linearly
independent representations of each dimension through linear transformation,
which can be used to extract the main feature components of data and is often
used for dimensionality reduction processing of high-dimensional data.
Principal component analysis (PCA) is often used to reduce dimensionally. PCA algorithm is the process of statistical calculation that summarizes data on a large data table through a small data set of "abstract index" with unchanged content. So that it can be easily imagined and analysand. The main idea of PCA is to make clear the pattern and correlation among various features in the data set, and after finding the strong correlation between different variables, reduce the data size in a way that still retains important data[34]. PCA is the linear feature learning method where the single vector has the linear function of data metrics. This single vector is produced through repetition and a simple algorithm. Such a process is critical for solving complex data-driven problems that involve the use of high-dimensional data sets.
1.5 Extended morphological profile（EMP）
Extended
Morphological Profiles in Hyperspectral imaging means that the datasets
available are more precise because of the integration with other datasets.
Morphological contour is constructed by morphological operation and has been
widely used in remote sensing image processing. For hundreds of hyperspectral
image bands. processing in each band is very inefficient and redundant.
Extended Morphological Profile (EMP) used linear divergence analysis (LDA) to
extract spectral features, and weighted adaptive filters (AWFs) to extract
spatial information. After multiple iterations, they were fused with LBP
features[35]. The experimental results proved that the method can extract
feature information further effectively. Therefore, introducing morphological
features and establishing them on multiple hyperspectral bands is a very
effective application method, but how to select representative images and
morphological operation is still a problem. In this case, Principal Component
Analysis (PCA) used EMP Differential Morphological Profiles (DMP) to extract
different features that achieved excellent performance. In the process of
morphological operation, there are operations such as "open
operation" and "closed operation", "corrosion" and
"expansion", and structural elements of different sizes need to be
defined before these operations[32].
0 মন্তব্য(গুলি):
একটি মন্তব্য পোস্ট করুন