INTERNATIONAL JOURNAL OF COMPUTERS COMMUNICATIONS & CONTROL
Online ISSN 1841-9844, ISSN-L 1841-9836, Volume: 17, Issue: 2, Month: April, Year: 2022
Article Number: 4541, https://doi.org/10.15837/ijccc.2022.2.4541

CCC Publications 

PLDANet: Reasonable Combination of PCA and LDA Convolutional
Networks

C. C. Zhang, M. Mei, Z. L. Mei, J. K. Zhang, A. Y. Deng, C. L. Lu

Caicai Zhang&
School of Modern Information Technology
Zhejiang Institute of Mechanical and Electrical Engineering
528 Binwen Road, Binjiang District, Hangzhou, 310053, Zhejiang, China
zhangcaicai@zime.edu.cn

Mei Mei&
Department of Ultrasound, The Second Affiliated Hospital
Zhejiang University School of Medicine
88 Jiefang Road, Shangcheng District, Hangzhou, 310009, Zhejiang, China
meimeizjdx@zju.edu.cn
&Co-first authors: these authors contributed equally to this work.

Zhuolin Mei
School of Computer and Big Data Science
Jiujiang University
551 Qianjin East Road, Jiujiang, Jiangxi, 332005, China
6120153@jju.edu.cn

Junkang Zhang
School of Computer and Big Data Science
Jiujiang University
No. 551, Qianjin East Road, Jiujiang, Jiangxi, 332005, China
6120144@jju.edu.cn

Anyuan Deng*
School of Computer and Big Data Science
Jiujiang University
No. 551, Qianjin East Road, Jiujiang, Jiangxi, 332005, China
dengay@jju.edu.cn
*Corresponding author: dengay@jju.edu.cn

Chenglang Lu
School of Modern Information Technology
Zhejiang Institute of Mechanical and Electrical Engineering
528 Binwen Road, Binjiang District, Hangzhou, 310053, Zhejiang, China
luchenglang@zime.edu.cn



https://doi.org/10.15837/ijccc.2022.2.4541 2

Abstract

Integrating deep learning with traditional machine learning methods is an intriguing research
direction. For example, PCANet and LDANet adopts Principal Component Analysis (PCA) and
Fisher Linear Discriminant Analysis (LDA) to learn convolutional kernels separately. It is not
reasonable to adopt LDA to learn filter kernels in each convolutional layer, local features of images
from different classes may be similar, such as background areas. Therefore, it is meaningful to
adopt LDA to learn filter kernels only when all the patches carry information from the whole
image. However, to our knowledge, there are no existing works that study how to combine PCA
and LDA to learn convolutional kernels to achieve the best performance. In this paper, we propose
the convolutional coverage theory. Furthermore, we propose the PLDANet model which adopts
PCA and LDA reasonably in different convolutional layers based on the coverage theory. The
experimental study has shown the effectiveness of the proposed PLDANet model.

Keywords: deep learning, principal component analysis, fisher principal component analysis,
convolutional coverage;

1 Introduction
Automatic image classification is an important task in many areas, such as face recognition [1, 15,

34], content-based image retrieval [2, 3, 29] and computer-aided medical images classification [4, 28, 30].
The performance of advanced automatic images classification models relys on good features abstracted
from images. Currently, deep convolutional neural networks, such as ConvNet [12, 13] and DenseNet
[10] are widely used for abstracting features from images, which use the famous back-propagation
algorithm to train all the parameters in the networks [20]. However, it lacks of intuitive interpretability,
theoretical analysis and results instability. Furthermore, the procedure of parameters training using
back-propagation algorithm is time-consuming, and its accuracy highly relies on hype-parameters
optimization [7, 27].

On the contrary, the traditional machine learning algorithms, such as principal components analysis
(PCA) [9] and linear dicriminant analysis (LDA) [23] have ready-by theoretics, and can calculate
perfect features of training data set directly. However, the time complexity of calculating eigen vectors
in PCA and LDA is O(n3), where n is the dimention of input data. The high time complexity prevents
them adopting in data sets with large dimention. For example, if the size of an image is 900pt∗700pt,
then the dimention of the input data is 630000. Therefore, some researchers consider to integrate
convolutional neural networks with principal component analysis or linear discriminant analysis. Chan
et al. [7] propose PCANet and LDANet. PCANet leverages unsupervised PCA to learn multi-stage
filters, while LDANet incorporates supervised LDA to learn convolutional kernels (or filters) for image
classification. Sun et al. [27] propose Fisher PCA (FPCA) to learn convolutional kernels based on
combination of LDA and PCA. However, to our knowledge, there are no existing works that study
how to combine PCA and LDA to learn convolutional kernels to achieve the best performance.

In these models, the labels of patches of images are the same as the labels of images when adopting
LDA in calculating each convolutional kernel. The patches of images with different categories may
have similar local features. For example, the patches of backgroud areas of different images may be
totally the same. Adopting LDA for calculating convolutional kernels may abstract unreasonable local
features, and harm the classification performance. Thus, although LDA takes advantage of additional
label information rather than PCA, it is suitable for adopting LDA to learn convolutional kernels only
when the patches of an image carry information from the whole image.

To our knowledge, there are no existing works that study how to combine PCA and LDA to
learn convolutional kernels to achieve the best performance. To abstract better features of images, we
propose an improved image classification model called PLDANet, based on a reasonable combination
way of PCA and LDA with deep convolutional neural networks.

Our main contributions in this article are summarized as follows.
(1) The reasonable combination schema of PCA and LDA is studied when integrating PCA and

LDA in calculating convolutional kernels. The combination schema is achieved based on a proposed
convolutional coverage theory, which calculates the number of necessary convolutional layers to garun-
tee that that each local patch of the image covers information from the whole image.



https://doi.org/10.15837/ijccc.2022.2.4541 3

(2) We propose a new model called PLDANet model on the basis of the theory of convolutional
coverage. PLDANet adopts a specific number of PCA convolutional layers before the subsequent LDA
convolutional layer. The specific number is calcuted according to the convolutional coverage theory,
so that the LDA convolutional layer takes advantage of labels information completely correctly.

(3) We conduct an extensive experimental study to evaluate PLDANet in different data sets. The
experimental results show that the proposed PLDANet have better performance on classification tasks.

The remainder of this article is organized as follows. Section 2 discusses the related work. Section
3 describes our proposed convolutional coverage theory. Section 4 describes our proposed PLDANet
model on the basis of the theory of convolutional coverage. In Section 5, we report the performance
evaluation of our techniques over five MNIST variations. Section 6 concludes this paper.

2 Related work
As big data containing images has been continuously generated from many areas[32, 33], automatic

image features extraction and classification method is needed in many applications. PCA produces
the most distinctive subspace for representation by maximizing the variance ,of the extracted features
and discarding noisy dimensions [9]. DPCA model [16] performs a two-layer zero-phase component
analysis (ZCA) whitening plus PCA structure to learn, hierarchical features for face recognition. The
whole feature representation of images is extracted by concatenating the representations from the two
layers.

To further improve the performance of traditional PCA based image classification, researchers have
proposed several versions of PCA in the past decade. Pereira el al. [22] proposes a modular PCA
algorithm that divides an image into smaller sub-images, which extracts local and global information
of images aiming to reduce the effects caused by variable changes. Liu et al. [17] combines the
traditional PCA and kernel-PCA to outline the data into high dimension vector space with non-linear
transformation.

To achieve a better computational and accuracy rate, some studies combine PCA with different
techniques, such as Delaunay Triangulation [14], Linear Discriminant Analysis [24], Deep Neural
Networks (DNN) [5] and Convolutional Neural Network (CNN) [21]. Machidon et al. [19] examine the
integration of PCA with Delaunay Triangulation. Jing et al. [11] integrate feature extraction of Color
2D-PCA, and Convolutional Neural Network into one dicision-level fusion. Generalized discriminant
analysis (GerDA) [26] is a generalization of the classical linear discriminant analysis, which uses DNN
to extract features. Lu et al. [18] propose a new joint feature learning approach to automatically
learn feature representations and stack a deep architecture to exploit hierarchical information. The
weighted PCA (WPCA)[31] is used in joint feature learning to map the combined outputs of the first
layer to a low-dimensional feature space for reducing the redundancy.

A CNN architecture consists of multiple trainable stages stacked on top of each other, followed by
a supervise classifier [25]. Chan et al. [7] propose two deep model variations, PCANet and LDANet.
PCANet integrates PCA with CNN and can easily be employed to learn multistage filters. PCANet
processes input images by cascaded PCA, binary hashing [6], and block histograms. PCANet is an
unsupervised deep learning baseline that mainly leverages PCA to learn multi-stage filters. While
LDANet is a supervised deep learning framework that mainly incorporates LDA with CNN to learn
filters. Sun et al. [27] propose Fisher PCA (FPCA) to learn each convolutional kernels based on
mixing of PCA and LDA. These models adopt LDA in calculating each convolutional kernel, and take
the label of an image as that of all patches of the image. However, the patches of images with different
categories may have similar local features. Unreasonable local features may be learned by adopting
LDA in calculating each convolutional kernel and lead to imperfect performance. Thus, it is suitable
for adopting LDA to learn convolutional kernels only when the patches of an image carry information
from the whole image. As far as we know, we are the first to study how to combine PCA and LDA
reasonably to learn convolutional filters.



https://doi.org/10.15837/ijccc.2022.2.4541 4

3 The convolutional coverage
In the LDA convolutional layer, the labels of all patches of a image are the same as the label of

the image. However, it is not reasonable to adopt LDA to learn filter kernels for every convolutional
layer, local features of images from different classes may be similar, such as background areas. It is
meaningful to adopt LDA to learn filter kernels only when all the patches carry information from the
whole image.

In this section, we propose a theory of convolutional coverage and analyze when a patch will
carry information from the whole image. Then a new convolutional network model based on the
convolutional coverage theory will be introduced in Section 4.

First, the definition of the convolutional coverage is given in Definition 1. The symbols in the
paper are summarized in Table 1.

Table 1: The summarization of symbols

Symbols Description

I ∈ Rm×n an image I with m ∗ n pixels

C[w, h] the convolutional coverage of the pixel [w, h] in I

Ci[w, h] the convolutional coverage of the pixel [w, h] in I in the i
th convolutional layer

[n] round a float point number to the next integer

Definition 1 (The convolutional coverage). Given an image I ∈ Rm×n and a pixel [w, h] in I, if
k1 × k2 is the kernel size of the convolutional layer, then we say that C[w, h] = [wmin : wmax, hmin :
hmax] is the convolutional coverage of the pixel [w, h] in I, where
C[w, h].wmin = max{0, w − k1/2},
C[w, h].wmax = min{m − 1, w + k1/2},
C[w, h].hmin = max{0, h − k2/2},
C[w, h].hmax = min{n − 1, h + k1/2}.

The convolutional coverage defines the range of the image carrying by a pixel after the convolutional
operation. For example, given a image with size 28 × 28, if the kernel size is 7 × 7, then C[6, 6] = [3 :
9, 3 : 9]

Then the method of computing the convolutional coverage of a pixel of an image in the ith convo-
lutional layer is given in Theorem 1.

Theorem 1. Suppose that I ∈ Rm×n is an image, I i ∈ Rm×n is the output in the ith convolutional
layer of I, and the kernel size is k1 ×k2 in each convolutional layer. If Ci[w, h] = [wmin : wmax, hmin :
hmax] is the ith convolutional coverage of the pixel [w, h] in I, then
wmin = max{0, w − i × k1/2},
wmax = min{m − 1, w + i × k1/2},
hmin = max{0, h − i × k2/2},
hmax = min{n − 1, h + i × k2/2}.

Proof. Let I 1 be the output in the first convolutional layer, then we have,
C1[w, h].wmin = max{0, w − k1/2},
C1[w, h].wmax = min{m − 1, w + k1/2},
C1[w, h].hmin = max{0, h − k2/2},
C1[w, h].hmax = min{n − 1, h + k1/2}.
Let I (i−1) ∈ Rm×n be the output in the (i−1)th convolutional layer, and C(i−1)[w, h] = [wmin : wmax, hmin :



https://doi.org/10.15837/ijccc.2022.2.4541 5

hmax] be the (i − 1)th convolutional coverage in I for the pixel [w, h], then we have,
Ci[w, h].wmin = max{0, C

(i−1)
[w−k1/2, h−k2/2].wmin}

= max{0, C(i−j)[w−j×k1/2, h−j×k2/2].wmin}
= max{0, C1[(w−(i−1)×k1/2, h−(i−1))×k2/2].wmin}
= max{0, w − (i − 1) × k1/2 − k1/2}
= max{0, w − i × k1/2}
Ci(w, h).wmax = min{m − 1, C(i−1)(w + k1/2, h + k2/2).wmax}
= min{m − 1, C(i−j)(w + j × k1/2, h + j × k2/2).wmax}
= min{m − 1, C1(w + (i − 1) × k1/2, h + (i − 1) × k2/2).wmax}
= min{m − 1, w + (i − 1) × k1/2 + k1/2}
= min{m − 1, w + i × k1/2}
Similarly,
Ci(w, h).hmin = max{0, h − i × k2/2},
Ci(w, h).hmax = min{n − 1, h + i × k2/2}

Example 1. Given an image with size 28×28, if the kernel size in each convolutional layer is 7×7,
then
C1[6, 6] = [3 : 9, 3 : 9]
C2[6, 6] = [0 : 12, 0 : 12]

Next we give the Theorem 2 to obtain the number of the convolutional layer that each patch of
the image carries information from the whole image.

Theorem 2. Given an image I ∈ Rm×n, and the kernel size is k1×k2 in each convolutional layer, after
G =[ max{(m − k1)/(k1/2), (n − k2)/(k2/2)}] convoluational layers, each patch carries information
from the whole image.

Proof. If the left upper patch [0 : (k1−1)), 0 : (k2−1))] covers information from the whole image, then
each patch does. If the pixel [(k1 − 1), (k2 − 1)] covers information from the pixel [(m − 1), (n − 1)],
then the left upper patch covers information from the whole image.
Based on Theorem 1,
CG[(k1−1), (k2−1)].wmax = min{(m − 1), k1 − 1 + G ∗ k1/2}
When k1 − 1 + G ∗ k1/2 = m − 1, we have
G = (m − k1)/(k1/2)
Similarly,
CG[(k1−1), (k2−1)].hmax = min{(n − 1), k2 − 1 + G ∗ k2/2}
When k2 − 1 + G ∗ k2/2 = n − 1, we have
G = (n − k2)/(k2/2)
Since G have to be an interger, we round up the value to the greater one of (m − k1)/(k1/2) and
(n − k2)/(k2/2).
Therefore, after G =[ max{(m−k1)/(k1/2), (n−k2)/(k2/2)}] convoluational layers, each patch carries
information from the whole image.

Example 2. Given an image with size 28×28, if the kernel size in each convolutional layer is 7×7,
then G =[ (28 − 7)/3] = 7.

4 The PLDANet model
The convolutional coverage theory calculates the necessary number of the convolutional layers,

which is important for designing reasonable combination of PCA and LDA convolutional networks.
Section 4.1 gives the definitions of PCA convolutional layer and LDA convolutional layer. Section 4.2
introduces the proposed PLDANet model.



https://doi.org/10.15837/ijccc.2022.2.4541 6

4.1 Definitions

PCA extracts main features to maximum the variance of data, which is an unsupervised learning
method. Definition 2 defines the PCA convolutional layer.

Definition 2 (PCA convolutional layer). A convolutional layer is called PCA Convolutional layer if
the filter kernel parameters are learned by PCA.

Let {Ii ∈ Rm×n×c}Ni=1 be the input images of the PCA convolutional layer, where N is the number
of input images, m × n × c is the size of each image Ii. A patch with the filter kernel size k1 × k2 is
taken around each pixel of each image. Let {Xi}Ni=1 ∈ Rk1k2×cmnN be the patch set.

The filter kernel parameters in a PCA convolutional layer are learned by PCA on the patch set.
Each eigen vector calculated by PCA corresponds to a filter kernel.

LDA extracts discriminative features to maximum the variance of data in different classes and
minimum the variance of data in the same class, which is a supervised learning method. Definition 2
defines the LDA convolutional layer.

Definition 3 (LDA convolutional layer). A convolutional layer is called LDA Convolutional layer if
the filter kernel parameters are learned by LDA.

The filter kernel parameters in a LDA convolutional layer is learned by LDA on the patch set, and
the labels of all patch of an image are the same as the label of the image. Each eigen vector calculated
by LDA corresponds to a filter kernel.

4.2 The model architecture

Figure 1 shows the detailed diagram of the proposed PLDANet model. First, we evaluate the
number of PCA convolutional layers in the PLDANet model. G = max{(m − k1)/(k1/2), (n −
k2)/(k2/2)}, where (m, n) is the size of the input image, and (k1, k2) is the size of each filter kernel.
Then the Deep PLDANet model is constructed by the number of G PCA convolutional layers and
a subsequent LDA convolutional layer. Then the Hashing, Histograms and the linear SVM stages
are the same as the PCANet. The training process of PLDANet for classification is demonstrated in
Algorithm 1.

Figure 1: The detailed diagram of the proposed PLDANet model

Example 3. Given an image with size 28×28, if the kernel size in each convolutional layer is 7×7,
then G =[ (28 − 7)/3] = 7.
In the PLDANet model, the convolutional network includes 7 PCA convolutional layers and a subse-
quent LDA convolutional layer.

5 Experiments
Since the training time of PCANet and LDANet is much less than that of CNN, and the superior

performance of PCANet and LDANet over CNN has been studied in [7, 27], the comparison of our
proposed PLDANet and CNN has not been conducted in our work. In this section, we compare
PLDANet with PCANet and LDANet.

We conduct experiments on MNIST and its four variations. The MNIST dataset is one of the
most popular datasets in image classification. It contains 60,000 training and 10,000 test images



https://doi.org/10.15837/ijccc.2022.2.4541 7

Algorithm 1 The training process of PLDANet for image classification
Input:

(1) The training image set {(Ii, yi)}Ni=1;
(2) Ii ∈ Rm×n;
(3) The size of filter kernel k1, k2;

Output: The results of classification.
1: (m, n) ← the size of the input image
2: G = max{(m − k1)/(k1/2), (n − k2)/(k2/2)}
3: for each l = 1 : G do
4: I li ← P CA_Convolutional_Layer(Ii), i = 1, 2, . . . , N
5: end for
6: I

(G+1)
i ← LDA_Convolutional_Layer(Ii), i = 1, 2, . . . , N

7: Hashing and Histograms stage: take I (G+1) as input;
8: Train linear SVM: take Histograms codes as input;
9: return The results of image classification.

of 10 hand-written digits. MNIST variations [8] introduce addition of rotation (rot), addition of a
background composed of random pixels (bg-rand) or patches extracted from a set of images (bg-img),
or combinations of these factors (rot-bg-img) to MNIST digit classification. These variations make
the problem become particularly challenging. The five classification datasets are summarized in Table
2. The size of all the images are 28 × 28. An instance of MNIST variations is shown in Figure 2.

Table 2: Details of the 5 classification tasks on MNIST variations.

Data Sets Description Train-Valid-Test
basic Smaller subset of MNIST 10,000-2000-10,000
rot MNIST with rotation 10,000-2000-50,000

bg-rand MNIST with noise background 10,000-2000-50,000
bg-img MNIST with image background 10,000-2000-50,000

bg-img-rot MNIST with rotation and image background 10000-2000-50000

(a) rot (b) bg-rand (c) bg-img (d) bg-img-rot

Figure 2: The number eight in the MNIST variations

To evaluate the performance of PCA convolutional layer and LDA convolutional layer in different
convolutional layers, we compare PCANet-i, PLDANet-i and PLDANet2-i, where PCANet-i means
there are number of i PCA convolutional layers, PLDANet-i means there are PCA convolutional layers
in the first (i-1) layers and LDA convolutional layer in the last layer, and PLDANet2-i means there
are PCA convolutional layers in the first 6 layers and LDA convolutional layer in the rest layers.

Limited to the memory size in our experimental environment, the hyperparameters of PCANet-i
and PLDANet-i are set as follows: (1) the number of filters are L1 = Li = 8, Lj = 1, 1 < j < i; (2)
the patch size is k1 = k2 = 7 in all convolutional layers; (3) the block size for local histograms in the



https://doi.org/10.15837/ijccc.2022.2.4541 8

output layer is set to be 7 × 7, and the block overlap ratio is 0.5.

Figure 3: Comparison of PCANet-i and
PLDANet-i on basic MNIST

Figure 4: Comparison of PCANet-i and
PLDANet-i on rot MNIST

We compare PCANet-i and PLDANet-i on five MNIST variations when varying i from 6 to 9.
Figure 3 shows the accuracy rates of PCANet-i and PLDANet-i on the basic MNIST. Figure 4 shows
the accuracy rates of PCANet-i and PLDANet-i on the rot MNIST. Figure 5 shows the accuracy rates
of PCANet-i and PLDANet-i on the bg-rand MNIST. Figure 6 shows the accuracy rates of PCANet-i
and PLDANet-i on the bg-img MNIST. Figure 7 shows the accuracy rates of PCANet-i and PLDANet-i
on the bg-img-rot MNIST.

Figure 5: Comparison of PCANet-i and
PLDANet-i on bg-rand MNIST

Figure 6: Comparison of PCANet-i and
PLDANet-i on bg-img MNIST

The experimental results show that (1) PLDANet-6 have the same accuracy rates as PCANet-6
on basic MNIST and bg-img MNIST, and have lower accuracy rates than PCANet-6 on bg-rand, rot
MNIST and rot-bg-img MNIST; (2) PLDANet-7 have higher accuracy rates than PCANet-7 on bg-img
MNIST, bg-rand, rot MNIST and rot-bg-img MNIST, and have the same accuracy rate as PCANet-6
on the basic MNIST; (3) PLDANet-i (i=8, 9) have higher accuracy rates than PCANet-i (i=8, 9) on
the five MNIST variations.

Based on Theorem 2, G = (m−k)/(k/2) = (28−7)/3 = 7. Therefore, each patch of the image will
cover information from the whole image after 7 convolutional layers. However, in the 6th convolutional
layer and 7th convolutional layer, some but not all pathes carry information from the whole image, and
LDA convolutional layer have better or worse performance than PCA convolutional layer. In the 8th
convolutional layer and 9th convolutional layer, all patches carry information from the whole image,
and LDA convolutional layer have better performance than the PCA convolutional layer.

We compare PCANet-i, PLDANet-i and PLDANet2-i on five MNIST variations when varying i
from 8 to 9. Figure 8 shows the accuracy rates of PCANet-i, PLDANet-i and PLDANet2-i on the
basic MNIST. Figure 9 shows the accuracy rates of PCANet-i, PLDANet-i and PLDANet2-i on the



https://doi.org/10.15837/ijccc.2022.2.4541 9

Figure 7: Comparison of PCANet-i and
PLDANet-i on bg-img-rot MNIST

Figure 8: Comparison of PCANet-i, PLDANet-i
and PLDANet2-i on basic MNIST

rot MNIST. Figure 10 shows the accuracy rates of PCANet-i, PLDANet-i and PLDANet2-i on the
bg-rand MNIST. Figure 11 shows the accuracy rates of PCANet-i, PLDANet-i and PLDANet2-i on
the bg-img MNIST. Figure 12 shows the accuracy rates of PCANet-i, PLDANet-i and PLDANet2-i
on the bg-img-rot MNIST.

Figure 9: Comparison of PCANet-i, PLDANet-i
and PLDANet2-i on rot MNIST

Figure 10: Comparison of PCANet-i, PLDANet-i
and PLDANet2-i on bg-rand MNIST

The experimental results show that PLDANet2-i (i=8, 9) have higher accuracy rates than PCANet-
i and PLDANet-i on all the five MNIST variations. Since PLDANet-7 have higher or the same accuracy
rates compared to PCANet-7 on all the five MNIST variations, LDA convolutional layer performs
better than PCA convolutional layer in the jth (j ≥ 7) convolutional layer. PLDANet2-9 performs
LDA convolutional layer during 7th, 8th and 9th convolutional layer. The phenomenon proves our
initial idea that adopting LDA convolutional layer is greatly beneficial for extracting features when
each patch of the image covers information from the whole image.

6 Conclusion
Adopting PCA in learning convolutional kernels takes much less time than training a CNN, and

avoids the laborious fine-tuning process. The supervision of the learned kernels can be further enhanced
by incorporating the labels information based on LDA. However, it is meaningful to assign class
labels to the patches and adopt LDA only when all the patches carry information from the whole
image. In this paper, we propose a convolutional coverage theory, which indicate that all patches



https://doi.org/10.15837/ijccc.2022.2.4541 10

Figure 11: Comparison of PCANet-i, PLDANet-i
and PLDANet2-i on bg-img MNIST

Figure 12: Comparison of PCANet-i, PLDANet-
i and PLDANet2-i on bg-img-rot MNIST varia-
tions

cover information from whole images after a specific number of convolutional layers. Based on the
convolutional coverage theory, we propose PLDANet model, which includes a specific number of
PCA convolutional layers and a subsequent LDA convolutional layer. Benefiting from the reasonable
combination of PCA and LDA, PLDANet enhances the representation ability of PCANet and LDANet.
Finally, a variety of experiments are conducted to demonstrate the effectiveness of the proposed
PLDANet.

Funding
This work was founded by the National Natural Science Foundation of China (No. 61962029); the

Natural Science Foundation of Jiangxi Province (No. 20202ACBL202005 and No. 20202BAB212006);
MOE (Ministry of Education in China), Youth Fund Project of Humanities and Social Sciences Re-
search (No.21YJCZH096); Science and Technology Fundation of Jiangxi Educational Committee (No.
GJJ201832).

Author contributions

Caicai Zhang and Mei Mei contributed significantly to the conception of the study and wrote the
manuscript; Zhuolin Mei contributed to analysis. Junkang Zhang performed the experiment; Anyuan
Deng and Chenglang Lu helped perform the analysis with constructive discussions.

Conflict of interest

The authors declare no conflict of interest.

References
[1] Ahmed, S.B.; Ali, S.F.; Ahmad, J.; Adnan, M.; and Fraz, M.M. (2020). On the frontiers of pose

invariant face recognition: a review. Artificial Intelligence Review, 53(4), 2571–2634, 2020.

[2] Alzyoud, F.Y.; Maqableh, W.; and Faiz Al, S. (2021). A semi smart adaptive approach for
trash classification. INTERNATIONAL JOURNAL OF COMPUTERS COMMUNICATIONS &
CONTROL, 16(4172), 1–13, 2021.



https://doi.org/10.15837/ijccc.2022.2.4541 11

[3] Arun, K.S.; Govindan, V.K.; and Kumar, S.D.M. (2020). Enhanced bag of visual words repre-
sentations for content based image retrieval: a comparative study. Artificial Intelligence Review,
53(3), 1615–1653, 2020.

[4] Cao, Z.; Duan, L.; Yang, G.; Yue, T.; and Chen, Q. (2019). An experimental study on breast
lesion detection and classification from ultrasound images using deep learning architectures. BMC
Medical Imaging, 19(1), 511–519, 2019.

[5] Carrara, F.; Falchi, F.; Caldelli, R.; Amato, G.; Fumarola, R.; and Becarelli, R. (2017). Detecting
adversarial example attacks to deep neural networks. 1–7, 2017.

[6] Carreira-Perpián, M.. and Raziperchikolaei, R. (2015). Hashing with binary autoencoders. In
2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.

[7] Chan, T.H.; Jia, K.; and Gao, S. (2015). Pcanet: A simple deep learning baseline for image
classification? IEEE TRANSACTIONS ON IMAGE PROCESSING, 24, 5017–5032, 2015.

[8] Dlv, S. and Sor, R. (2007). An empirical evaluation of deep architectures on problems with many
factors of variation. In 24th International Conference on Machine Learning. 473–480, 2007.

[9] Groth, D.; Hartmann, S.; Klie, S.; and Selbig, J. (2013). Principal components analysis. Methods
Mol Biol, 930, 527–547, 2013.

[10] Huang, G.; Liu, Z.; Pleiss, G.; Maaten, L.V.D.; and Weinberger, K. (2019). Convolutional net-
works with dense connectivity. IEEE transactions on pattern analysis and machine intelligence,
1(4), 1–12, 2019.

[11] Jing, L.; Tao, Q.; Chang, W.; Kai, X.; and Fang-Qing, W. (2018). Robust face recognition using
the deep c2d-cnn model based on decision-level fusion. Sensors, 18(7), 2080, 2018.

[12] Krizhevsky, A.; Sutskever, I.; and Hinton, G.E. (2017). Imagenet classification with deep convo-
lutional neural networks. Communications of the ACM, 60(6), 84–90, 2017.

[13] Lecun, Y. and Bottou, L. (1998). Gradient-based learning applied to document recognition.
Proceedings of the IEEE, 86(11), 2278–2324, 1998.

[14] Lee, D.T. and Lin, A.K. (1986). Generalized delaunay triangulation for planar graphs. Discrete
& Computational Geometry, 1, 201–217, 1986.

[15] Lintang, R.A.; Purnawarman, M.; and Wibowo, E.P. (2017). Human face recognition application
using pca and eigenface approach. In 2017 Second International Conference on Informatics and
Computing (ICIC), 2017.

[16] Liong, V.E.; Lu, J.; and Wang, G. (2013). Face recognition using deep pca. In 2013 9th Interna-
tional Conference on Information, Communications & Signal Processing (ICICS). 1–5, 2013.

[17] Liu, C.; Zhang, T.; Ding, D.; and Lv, C. (2016). Design and application of compound kernel-pca
algorithm in face recognition. In 2016 35th Chinese Control Conference (CCC). 4122–4126, 2016.

[18] Lu, J.; Liong, V.E.; Wang, G.; and Moulin, P. (2017). Joint feature learning for face recognition.
IEEE Transactions on Information Forensics and Security, 10(7), 1371–1383, 2017.

[19] Machidon, A.L.; Machidon, O.M.; and Ogrutan, P.L. (2019). Face recognition using eigenfaces,
geometrical pca approximation and neural networks. In 2019 42nd International Conference on
Telecommunications and Signal Processing (TSP), 2019.

[20] Ooyen, A.V. and Nienhuis, B. (1992). Improving the convergence of the back-propagation algo-
rithm. Neural Networks, 5, 465–471, 1992.



https://doi.org/10.15837/ijccc.2022.2.4541 12

[21] Oziuddeen, M.A.K.; Poruran, S.; and Caffiyar, M.Y. (2020). A novel deep convolutional neu-
ral network architecture based on transfer learning for handwritten urdu character recognition.
Tehnički vjesnik, 27(4), 1160–1165, 2020.

[22] Pereira, J.F.; Barreto, R.M.; Cavalcanti, G.D.C.; and Tsang, I.R. (2011). A robust feature extrac-
tion algorithm based on class-modular image principal component analysis for face verification.
In 2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP).
1469–1472, 2011.

[23] Riffenburgh, R. and Clunies-Ross, C. (1960). Linear discriminant analysis. Pacific Science, 14,
251–256, 1960.

[24] Riffenburgh, R.H. and Clunies-Ross, C.W. (2013). Linear discriminant analysis. Pacific Science,
3(6), 27–33, 2013.

[25] Shin, H.C.; Roth, H.R.; Gao, M.; Lu, L.; Xu, Z.; Nogues, I.; Yao, J.; Mollura, D.; and Summers,
R.M. (2016). Deep convolutional neural networks for computer-aided detection: Cnn architec-
tures, dataset characteristics and transfer learning. IEEE Transactions on Medical Imaging, 35(5),
1285–1298, 2016.

[26] Stuhlsatz, A.; Lippel, J.; and Zielke, T. (2014). Feature extraction with deep neural networks
by a generalized discriminant analysis. IEEE Transactions on Neural Networks and Learning
Systems, 23(4), 596–608, 2014.

[27] Sun, K.; Zhang, J.; Yong, H.; and Liu, J. (2018). Fpcanet: Fisher discrimination for principal
component analysis network. Knowledge-Based Systems, 166, 108–117, 2018.

[28] Swati, Z.N.K.; Zhao, Q.; Kabir, M.; Ali, F.; and Lu, J. (2019). Brain tumor classification for
mr images using transfer learning and fine-tuning. Computerized Medical Imaging and Graphics,
75(7), 34–46, 2019.

[29] Unar, S.; Wang, X.; Wang, C.; and Wang, Y. (2019). A decisive content based image retrieval
approach for feature fusion in visual and textual images. Knowledge-Based Systems, 179, 8–20,
2019.

[30] Xiaofeng; Qi; Lei; Zhang; Yao; Chen; Yong; Pi; Yi; and and, Q. (2018). Automated diagnosis of
breast ultrasonography images using deep neural networks. Medical Image Analysis, 52, 185–198,
2018.

[31] Z., L.; M., P.; and S.Z., L. (2014). Learning discriminant face descriptor for face recognition.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(2), 289–302, 2014.

[32] Zhang, C.; Mei, Z.; Wu, B.; Yu, J.; and Wang, Q. (2020). Query with assumptions for probabilistic
relational databases. Tehnički vjesnik, 27(3), 923–932, 2020.

[33] Zhang, S.; Yang, L.T.; Feng, J.; Wei, W.; Cui, Z.; Xie, X.; and Yan, P. (2021). A tensor-network-
based big data fusion framework for cyber-physical-social systems (cpss). Information Fusion,
(76), 337–354, 2021.

[34] Zhou, Y.; Wang, Y.; and Wang, X.H. (2018). Face recognition algorithm based on wavelet
transform and local linear embedding. Cluster Computing, 22, 1529–1540, 2018.



https://doi.org/10.15837/ijccc.2022.2.4541 13

Copyright ©2022 by the authors. Licensee Agora University, Oradea, Romania.
This is an open access article distributed under the terms and conditions of the Creative Commons
Attribution-NonCommercial 4.0 International License.
Journal’s webpage: http://univagora.ro/jour/index.php/ijccc/

This journal is a member of, and subscribes to the principles of,
the Committee on Publication Ethics (COPE).

https://publicationethics.org/members/international-journal-computers-communications-and-control

Cite this paper as:
Zhang C.C.; Mei, M.; Mei, Z.L.; Zhang, J.K.; Deng A.Y.; Lu, C.L. (2022). PLDANet: Reasonable
Combination of PCA and LDA Convolutional Networks, International Journal of Computers Com-
munications & Control, 17(2), 4541, 2022.

https://doi.org/10.15837/ijccc.2022.2.4541


	Introduction
	Related work
	The convolutional coverage
	The PLDANet model
	Definitions
	The model architecture

	Experiments
	Conclusion