over-sampling methods, namely border-SMOTE1 and border-SMOTE2 were given,
and the method gets better True Positive (TP) rate and F-value. Although the data
processing level has achieved some satisfactory results for imbalanced classification,
they still have their drawbacks. It is easy to lose significant information or add trivial
information, so as to affect the recognition accuracy of the minority.
At the same time, many machine learning methods have been focused on detailed
techniques to cope with imbalanced data classification. Datta et al. [9] proposed the
approach called near-Bayesian support vector machine (NBSVM) to adapt for cases
with the skewed distribution misclassification. Support vector machine (SVM), as the
basic classifiers of the ensemble committee, improved the accuracy of the hyperspectral
remote sensing images imbalanced classification [10]. Yan et al. [5] used an extended
bootstrapping and convolutional neural networks method to handle the skewed multimedia datasets. The results showed that the framework can work effectively and
greatly reduce the running time. These papers provide us with a new idea of whether
shallow models can be turned into depth architectures.
In this paper, the deep support vector machine (DSVM) algorithm has shown
excellent performance in regression, classification and dimension reduction [11, 12].
Hence, we took into account the DSVM algorithm for imbalanced data classification.
To the best of our knowledge, the DSVM algorithm is the first time to be applied to
imbalanced human target detection. Furthermore, we fine-tune the DSVM algorithm to
obtain promising performance of imbalanced human target detection.
The rest of this paper is organized as follows. Section 2 introduces the DSVM
algorithm. In Sect. 3, we describe the experimental process and results in detail. It is
also compared with SVM and the SAE algorithm. Numerical experimental results show
that the DSVM algorithm performs better. In the end, Sect. 4 summarizes this paper
and points out possible directions for future work.
2 The Deep Support Vector Machine Algorithm
Wiering et al. [11] proposed the DSVM algorithm, which replaces neurons as SVM and
has depth models. The DSVM algorithm is the application of the deep learning model
to SVM. First, it trains SVM in a standard way. Secondly, the kernel activation
function of the support vector is used as input data into the SVMs of the hidden layer.
Finally, the main SVM is trained to establish a nonlinear combination of the kernel
activation function of the support vector. The module of a two-layer DSVM is shown
in Fig. 1.
2.1 The DSVM Algorithm Model
We choose a classification dataset: x n 2 R
m
; y n 2 R
s
f
g
N
n¼1 , where x n is input vector and
y n is the scalar target output. The connection between the input vector x n and the output
target y n is
88
L. Zhang et al.
and the method gets better True Positive (TP) rate and F-value. Although the data
processing level has achieved some satisfactory results for imbalanced classification,
they still have their drawbacks. It is easy to lose significant information or add trivial
information, so as to affect the recognition accuracy of the minority.
At the same time, many machine learning methods have been focused on detailed
techniques to cope with imbalanced data classification. Datta et al. [9] proposed the
approach called near-Bayesian support vector machine (NBSVM) to adapt for cases
with the skewed distribution misclassification. Support vector machine (SVM), as the
basic classifiers of the ensemble committee, improved the accuracy of the hyperspectral
remote sensing images imbalanced classification [10]. Yan et al. [5] used an extended
bootstrapping and convolutional neural networks method to handle the skewed multimedia datasets. The results showed that the framework can work effectively and
greatly reduce the running time. These papers provide us with a new idea of whether
shallow models can be turned into depth architectures.
In this paper, the deep support vector machine (DSVM) algorithm has shown
excellent performance in regression, classification and dimension reduction [11, 12].
Hence, we took into account the DSVM algorithm for imbalanced data classification.
To the best of our knowledge, the DSVM algorithm is the first time to be applied to
imbalanced human target detection. Furthermore, we fine-tune the DSVM algorithm to
obtain promising performance of imbalanced human target detection.
The rest of this paper is organized as follows. Section 2 introduces the DSVM
algorithm. In Sect. 3, we describe the experimental process and results in detail. It is
also compared with SVM and the SAE algorithm. Numerical experimental results show
that the DSVM algorithm performs better. In the end, Sect. 4 summarizes this paper
and points out possible directions for future work.
2 The Deep Support Vector Machine Algorithm
Wiering et al. [11] proposed the DSVM algorithm, which replaces neurons as SVM and
has depth models. The DSVM algorithm is the application of the deep learning model
to SVM. First, it trains SVM in a standard way. Secondly, the kernel activation
function of the support vector is used as input data into the SVMs of the hidden layer.
Finally, the main SVM is trained to establish a nonlinear combination of the kernel
activation function of the support vector. The module of a two-layer DSVM is shown
in Fig. 1.
2.1 The DSVM Algorithm Model
We choose a classification dataset: x n 2 R
m
; y n 2 R
s
f
g
N
n¼1 , where x n is input vector and
y n is the scalar target output. The connection between the input vector x n and the output
target y n is
88
L. Zhang et al.