Hybrid Shallow Learning and Deep Learning for Feature Extraction
and Image Retrieval
Hanen Karamti
1,2 a
, Hadil Shaiba
1b
and Abeer M.Mahmoud
1,3 c
1
Computer Sciences Department, College of Computer and Information Sciences,
Princess Nourah bint Abdulrahman University, PO Box 84428, Riyadh, Saudi Arabia
2
MIRACL Laboratory, ISIMS, University of Sfax, B.P. 242, 3021 Sakiet Ezzit, Sfax, Tunisia
3
Faculty of Computer and Information Sciences, Ain Shams University, Cairo, Egypt
Keywords: Deep Convolutional Neural Network, Image Retrieval, Fuzzy C-Means Clustering, Feature Extraction.
Abstract: In the last decennia, several works have been developed to extract global/or local features from images.
However, the performance of image retrieval stay surfing from the problem of semantic interpretation of the
visual content of images (semantic gap). Recently, deep neural networks (DCNNs) showed excellent
performance in different fields like image retrieval for feature extraction compared to traditional techniques.
Although, Fuzzy C-Means (FCM) Clustering Algorithm that is a shallow learning method, but it has a
competitive performance in the clustering field. In this paper, we present a new method for feature extraction
combining DCNN and Fuzzy c-means, where DCNN gives a compact representation of images and FCM
clusters the features and enhances the real-time for searching. The proposed method is performed against
other methods in literature on two benchmark datasets: Oxford5K and Inria Holidays, where the proposed
method overbeats respectively 83% and 86%.
1 INTRODUCTION
Every day, many numerical images are born in
several models like medicine, science, and biology.
This big mass needs efficient tools for indexing,
searching and retrieving images like the CBIR
(Content-Based Image retrieval) systems. CBIR has
been a hot research topic in computer vision. It
searches images from datasets based on their visual
content. Visual content means low-level features,
called also visual features, extracted from images by
local and/or global descriptors. Global descriptors
(Varish et al., 2015, Nazir et al., 2018) extract from
images color, texture and Shape (Yuan et al., 2011,
Hiremath et al., 2007, Yu et al. 2013). Color (Gopa et
al., 2015) represents the most popular unit to describe
the visual content of images. Color histograms, Color
layout, and scalable color are the most used
descriptors. The texture is the homogeneity surface
quality of the object. To extract textures from images,
a
https://orcid.org/0000-0001-5162-2692
b
https://orcid.org/0000-0003-1652-6579
c
https://orcid.org/0000-0002-0362-0059
several descriptors are proposed like the statistical
methods, spectral (Qin et al., 2015).
The shape (Wu et al., 2009) is the two-
dimensional object contoured by a line. The basic
geometric types of shapes are oval, square, triangle
circle and rectangle. Several shape descriptors are
proposed like Fourier descriptors, moment invariants
(Varish et al., 2015, Hiremath et al., 2007).
Local descriptors (Wang et al., 2007) are based on
the low-level description of images but focused on a
specific area or region in it. The most popular
descriptors in this category are FAST (Features from
accelerated segment test), Harris corner detectors,
points of interest (POI) detectors, Scale-invariant
feature transform (SIFT) (Sun et al., 2017, Yuan et
al., 2011), blob detectors, Speeded Up Robust
Features (SURF) (Sun et al., 2017) and Locally
Aggregated Descriptors (VLAD) (Jégou et al., 2010).
To calculate the similarity between images, we
compare their low-level features. Several measures
Karamti, H., Shaiba, H. and Mahmoud, A.
Hybrid Shallow Learning and Deep Learning for Feature Extraction and Image Retrieval.
DOI: 10.5220/0009417501650172
In Proceedings of the 22nd International Conference on Enterprise Information Systems (ICEIS 2020) - Volume 1, pages 165-172
ISBN: 978-989-758-423-7
Copyright
c
2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
165
are used like Euclidian, Manhattan, and Chi2 (Zhou t
al, 2019).
Many CBIR systems (Paulin et al., 2015, Piras et
al., 2010) were developed to extract the visual
features from images. The traditional techniques were
developed around the use of global or local
descriptors and the use of similarity measures. Other
systems were developed to enhance retrieval
performance by the use of machine learning
techniques likes SVM (Fu et al., 2016), k-means
(Karamti et al., 2013), neural network (Karamti et al.,
2018), etc.
Recently, deep learning (Nguyen et al., 2018,
Gong et al., 2014, Qassim et al., 2018, Howard et al.,
2017, Vincent et al., 2010) achieves an important
success in different domains, especially in image
retrieval. Therefore, this new field encourages many
researchers to integrate it into their researches to
enhance clustering, recognition and classification
tasks.
In (Razavian et al., 2014), they developed three
models of CNN for feature extraction: basic CNN,
CNN with a linear SVM and CNN with additional
information including the cropped and rotated
examples. Radenovic et al. (Radenovic et al., 2017)
proposed a new model for image retrieval called
Generalized-Mean (GeM) that learned a 3D model
and employed the PCA to improve the vector of
features. Authors in (Mohedano et al., 2016), built a
simple pipeline for retrieving that employs CNN with
encoded convolutional features combined with the
bag of words (BoW). Arandjelovic et al.
(Arandjelovic et al., 2016) developed a new method
NetVLAD that employs a new CNN architecture
based on VLAD for image recognition.
In (Gordo et al., 2016), authors represented a
CNN model fully connected that uses many
convolutional layers combined with global
descriptors after features reduction. Authors in (Qin
et al., 2015) added a VLAD in CNN in the last layer
of model, which performed excellent results in image
retrieval. In (Qin et al., 2015, Razavian et al., 2016)
authors proposed other methods for deep learning
improvement as the feature extraction using Neural
codes (Babenko et al., 2014) or Spatial pooling
(Razavian et al., 2016).
In (Fu et al., 2016), the authors used a hybrid
model based on CNN and a support vector machine
(SVM). In (Ghrabat et al., 2019), authors extracted
global features like texture (co-occurrence matrix)
and they clustered features using the technique. Then,
they applied the GA (genetic algorithm) to classify
features by the use of SVM. The proposed deep
model is CNN.
Recently, several deep learning techniques are
proposed but DCNN was recommended as the best
enhancement in image retrieval (Babenko et al., 2015,
Kim et al., 2018, Lin et al., 2018). This enhancement
concerns the reduction of computational cost, which
is the target of many researchers. DCNN learns
several features automatically with consideration of
the semantic gap factor (Babenko et al., 2015, Kim et
al., 2018, Lin et al., 2018, Gordo et al., 2016). In
addition, deep learning mechanism can handle a large
and different collection of images that represents a
good performance in several tasks of image retrieval
(He, et al., 2014). In addition, DCNN (Wu et al.,
2018, Wu et al., 2018) was used for text classification,
and the obtained results are performant again other
methods that use the shallow techniques for
classification.
Figure 1: Feature extraction and retrieving process.
ICEIS 2020 - 22nd International Conference on Enterprise Information Systems
166
In this paper, we propose a new CBIR system that
profits from the DCNN and FCM advantages. A set
of experiments were presented to test the proposed
method using two benchmarks: Oxford5K and Inria
Holidays, where obtained results displayed promising
results. The remaining paper is organized as follows:
section 2 describes the proposed method. Section 3
discusses the results. Finally, section 4 presents the
conclusion and further research directions.
2 CONTENT-BASED IMAGE
RETRIEVAL USING DCNN
AND FUZZY C-MEANS
In this paper, we profit from the benefits of DCNN
and FCM to minimize the sharing time. The use of
DCNN replace the visual descriptors to extract the
low-level features. This convolutional model is a
deep representation, composed by a set of nonlinear
operations belonging to multiple levels. FCM is used
to enhance accelerate the retrieving process.
Our approach, represented by Figure 1, contains
three main phases: the first phase concerns features
extraction. Second phase is clustering and the third
phase is similarity calculation. The first phase takes
place offline where features are extracted from
images using the DCNN model. Then, the FCM
technique is applied in the second phase to cluster
images into similar clusters. After that, a label
signature is assigned to each cluster to distinguish
between clusters. All the labels reconstruct the output
layer of DCNN that serves in retrieving phase. Last
phase represents the retrieving phase that calculates
similarity between the features extracted from query
and label signatures. Based on the similarity values,
the closest class is returned containing the relevant
images.
2.1 Deep Neural Network
To extract the low-level features from images, we
adopted the Deep Convolutional Neural Network
architecture (DCNN). Several researches in the image
retrieval field focused on the use of DCNN to extract
the features because this algorithm is able to produce
better results despite its high computational cost.
DCNN is a Deep Neural Network with additional
convolutional layers. It has an architecture lighter
than other deep learning models. A large network
having multiple convolutional layers presents this
architecture. Figure 2 displays the proposed network
that contains 7 layers. The first layer represents the
input data. The second layer is a convolutional layer
of size 256*256*3. Both Layer 3 and 5 are pooling
layers using the function Softmax pooling w ith filter
3*3. The fourth layer is a convolutional layer with
size 64*64*3. Layer 6 represents a fully connection
layer having 4096 neurones. The last Layer (layer 7)
is the output layer containing 1024 neurones where
each neuron presents a low-level feature. For feature
extraction, the values of last fully connected layer in
DCNN model, represented but the output layer in
Figure 2, is excluded from the network and combined
into one vector of features. The dimension of this
vector is 1024 that will be used to search the similar
images. The feature extraction process is done offline
to extract features from images and online to extract
features from the query.
Figure 2: DCNN architecture.
2.2 Fuzzy C-Means Clustering
Algorithm
After feature extraction, images from the dataset are
clustered using a shallow learning technique “fuzzy
c-means” that is given by Figure 3.
Fuzzy c-means is one of the most popular
algorithm used in different area of research including
Hybrid Shallow Learning and Deep Learning for Feature Extraction and Image Retrieval
167
computer science. It is the fuzzy version of k-means
algorithm. FCM is used to make data clustering by a
set of features and a number of initialized clusters. Let
l is number of clusters randomly initialized. The
target of FCM is to cluster the similar genes in a
single cluster. Gene seems the vector of features. Let
Xx
:i1,…,n, where x
is the value of gene g
.
The goal is to cluster all the gene into c clusters
where c∈2…n 1. The cluster c
is assigned by a
partition matrix Ww
,
that contain the degree of
membership of gene g
in cluster c
.
For a cluster c
, the corresponding centroid
Cent
is defined as:
Cent
∑
∑
(1)
Where p ∈ [1..∞[ is a parameter that determines
the influence of the weights w
,
.
Figure 3: FCM algorithm.
The distance between a centroid (Cent
) and a
gene g
represented by the features vector x
, is
calculated using Euclidean distance, as the following
equation where z is the size of g
:
distx
,Cent
∑
x
,Cent
(2)
FCM attempts to minimize the cost function
designer by the sum of the squared error (SSE), as
with k-means algorithm.
f
∑∑
w
,
distx
,Cent
(3)
After clustering, a label signature is assigned for
each cluster. We means by label signature the
signature of the centroid that is represented by vector
of features.
2.3 Similarity Calculation
To search for images based on their low-level
features, Euclidian distance (see equation 4, where k
is the number of features and j∈0,c]) is used.
Similarity is calculated between the features of query
and the label signatures of images. This phase is done
online and all the obtained similarity scores are sorted
in a descending order. The relevant label signature
should have the first rank. Then the relevant cluster
corresponding to its label is returned, and all the
images belonging to that cluster are considered
relevant.
Simq,Cent
∑
q,Cent
(4)
On the other hand, images in the relevant cluster
can be displayed in ascending order. In this case, a
similarity score should be calculated between the
label signature and the cluster images.
3 EXPERIMENTS
In this section, our contribution is evaluated using a
set of experiments: the first set to evaluate the
parameters of DCNN and the second set is proposed
to compare the developed method with other state-of-
the-art methods. The experiments are performed
using two datasets:
INRIA Holidays dataset contains personal
holiday’s photos taken in different scene types.
Images of each scene regroups almost 500 images.
The first image (image number one) from each scene
represents the query image.
Oxford5K dataset contains 5062 images selected
from Flickr displayed a specific Oxford landmark. 11
different landmarks are regrouped as Ground truth
where each one is represented by five possible
queries. Therefore, number of queries equals 55.
Performance is evaluated using the mean Average
Precision (mAP) function.
ICEIS 2020 - 22nd International Conference on Enterprise Information Systems
168
Figure 4: Results on Oxford with FCM.
Figure 5: Results on Oxford without FCM.
Figure 6: Results on Holidays with FCM.
Figure 7: Resultson Holidays without FCM.
3.1 Evaluation of DCNN for Feature
Extraction and FCM for Clustering
In this section, we evaluate the proposed network,
especially the DCNN and the FCM models.
Concerning DCNN, we test six version of the
proposed model. Each version includes different
parameters including convolutional layer, pooling
layers, fully connected layers and the activation
function. In addition, each version contains 7 layers:
input layer, 2 convolutional layers, 2 pooling layers,
1 layer full connection, and output layer. The six
versions are:
DCNN1 stars with a convolutional layer having the
size 224*224*3 and finishes with size equals to
14*14*512. The pooling layers include an average
pooling function.
DCNN2 has identical set of parameters as DCNN1,
but the pooling layers uses a Softmax pooling
function.
DCNN3 begins with a convolutional layer of size
260*260*3 and output size equals to 48*48*512. The
pooling layers include an average pooling function.
DCNN4 has the same parameters that DCNN3, but
the pooling layers uses a Softmax pooling function.
DCNN5 is the proposed DCNN model described in
the previous section but uses an average pooling
function instead of Softmax pooling.
DCNN6 is the proposed DCNN model described in
section 2.1.
We evaluate each version separately and we re-
evaluate them with the proposed FCM model. The
evaluation is done on two steps. First step concerns
the search of queries using the features extracted from
each model in an isolation mode. In second step
evaluated the combination of that model with FCM.
Figures 4 and 5 display respectively the results on
Oxford dataset with and without FCM and figures 6
and 7 represent respectively the results on Holidays
dataset with and without FCM. The obtained results
show that DCNN6 is the best model compared with
other models containing or no FCM model. For the
rest of paper, we complete experiments using
DCNN6 model.
Table 1 represents the obtained results using
Oxford and Holidays datasets using DCNN combined
with or without FCM. On Oxford5k, the mAP
performed 0.83 with the use of FCM against 0.79
without FCM. However, the values of mAP on
Holidays are respectively 0.86 for DCNN and FCM
Hybrid Shallow Learning and Deep Learning for Feature Extraction and Image Retrieval
169
and 0.78 using DCNN only. Therefore, we conclude
that the use of FCM enhances the proposed model.
Table 1: Impact of FCM on Oxford and Holidays dataset
when it is combined with DCNN6.
Method Oxford5K Holidays
DCNN+FCM 0.83 0.86
DCNN 0.79 0.78
3.2 Similarity Measure Evaluation
The choice of the similarity measure is an important
step in a CBIR system. Therefore, to calculate the
similarity between the label signatures and query we
evaluated three distance measures: Chebyshev
(equation 5), Manhattan (equation 6) and Euclidian
(equation 4).
Chebyshevq,Cent
maxqCent
(5)
Manhattanq,Cent
∑
q Cent
(6)
Table 2 concludes the obtained results on Oxford
and Holidays using the three measures. From this
table, we conclude that Euclidian distance gives the
best results against the other distance measures. Is not
the first time to show the performance of Euclidian
distance in image retrieval. This measure is easy to
use and very fast compared to other measures that
need more computation time
Table 2: Evaluation of three similarity measures.
Measure Oxford5K Holidays
Chebyshev 0.79 0.8
Manhattan 0.77 0.82
Euclidian 0.83 0.86
3.3 (DCNN+FCM) versus Literatures
In this section, we compare the proposed model
(DCNN and FCM) with other methods from literature
that are recently developed. Figure 8, displays the
obtained results for testing different deep learning
models on Oxford5k and Holidays collections.
Using a basic Convolutional Neural Network
(CNN) (Razavian et al., 2014), we obtained on
Oxford mAP=32.2 and on Holidays, we obtained
mAP=64.2. Changing the pooling function from
average to max pooling (Razavian et al., 2016), the
MAP can reach respectively 53.3 on Oxford and 71.6
on Holidays. NetVLAD (Arandjelovic et al.,
2016) is the system that combines CNN and
VLAD, the mAP achieves 71.6 on Oxford5k and 87.5
on Holidays. The local descriptors like the bag of
Word (Mohedano et al., 2016) can be merged to CNN
to enhance the retrieval performance. Therefore, the
mAP achieves 73.9 on Oxford. The model proposed
by (Radenovic et al., 2017) (CNN-Gem) achieves the
best results compared to the above methods and our
method. The results are 87.8 on Oxford and 93.9 on
Holidays.
Concerning our method, the obtained results are
better than (Razavian et al., 2014, Razavian et al.,
2016, Mohedano et al., 2016). This shows that the
architecture of DCNN is well-suited feature
extraction compared to CNN. In addition, we
compared our results with (Arandjelovic et al., 2016)
and (Radenovic et al., 2017) that have result more
than we on the two datasets have. This difference is
around 3% on Oxford and between 1% and 6% on
Holidays. As recorded the difference between the
proposed method and literatures is relatively small,
however the proposed method over beats in the
number of used parameters. It means, similar
results are obtained with a lower cost and minimized
Figure 8: Comparison between our method and other methods from literature.
ICEIS 2020 - 22nd International Conference on Enterprise Information Systems
170
searching time (Arandjelovic et al., 2016, Radenovic
et al., 2017) despite the fact that they have important
values of mAP. In addition, it is not a benefit for such
methods to be costly when we talk about real-time
retrieving.
4 CONCLUSION AND FUTURE
WORK
In this work, a new method for image retrieval based
on the visual content of images was proposed. This
method uses the DCNN technique for feature
extraction. Then, it clusters the dataset and gives a
label signature for each cluster. Finally, the similarity
is calculated between the query and the labels to
accelerate the retrieving process.
Performance was evaluated using two datasets
Oxford5k and Holidays. The obtained results
displayed the efficiency of the proposed method.
Especially, when it was compared with other CBIR
systems on from literature.
In future work, retrieving performance can be
improved by the use of recent deep learning
techniques like the Generative adversarial network
(GAN).
REFERENCES
Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.,
2016. NetVLAD: CNN architecture for weakly
supervised place recognition, in CVPR.
Babenko, A., Lempitsky, V., 2015. Aggregating local deep
features for image retrieval, in The IEEE International
Conference on Computer Vision (ICCV).
Babenko, A., Slesarev, A., Chigorin, A., Lempitsky, V.,
2014. Neural codes for image retrieval. In: European
conference on computer vision, Springer pp.584–599.
Fu, R., Li, B., Gao, Y., Wang, P., 2016. Content-based
image retrieval based on CNN and SVM. In: 2
nd
IEEE
Int. Conf. on Computer and Communications (ICCC),
pp.638–642.
Ghrabat, M.J.J., et al., 2019. An effective image retrieval
based on optimized genetic algorithm utilized a novel
SVM-based convolutional neural network classifier.
Hum. Cent. Comput. Inf. Sci. 9, 31.
Gong, G., Wang, L., Guo, R., Lazebnik, S., 2014. Multi-
scale orderless pooling of deep convolutional activation
features. Proceedings of the European Conference on
Computer Vision, pp.392-407.
Gopal, N., SBhooshan, R., 2015. Content based image
retrieval using enhanced surf. In Fifth National
Conference on Computer Vision, Pattern Recognition,
Image Processing and Graphics (NCVPRIPG), pp.1-4.
Gordo, A., Almazan, A., Revaud, J., Larlus, D., 2016. Deep
image retrieval: Learning global representations for
image search. In European Conference on Computer
Vision, pp.241-257.
He, K., Zhang, X., Ren, S., Sun, J., 2014. Spatial Pyramid
Pooling in Deep Convolutional Networks for Visual
Recognition. IEEE Transactions on Pattern Analysis
and Machine. vol.37, pp.1904-1916.
Hiremath, P.S., Pujari, J., 2007. Content based image
retrieval using color, texture and shape features.
International conference on advanced computing and
communications. Guwahati, Assam, pp.780-784.
Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang,
W., Weyand, T., Andreetto, M., Adam. H., 2017.
MobileNets: Efficient Convolutional Neural Networks
for Mobile Vision Applications.
Jégou, H., Douze, M., Schmid, C., Pérez, P., 2010.
Aggregating local descriptors into a compact image
representation. In Computer Vision and Pattern
Recognition (CVPR), pp.3304-3311.
Karamti, H., 2013. Vectorisation du modèle d'appariement
pour la recherche d'images par le contenu. Conference
en Recherche d'Infomations et Applications (CORIA),
pp.335-340.
Karamti, H., et al., 2018. Vector space model adaptation
and pseudo relevance feedback for content-based image
retrieval. Multimedia Tools and Applications. vol.77,
pp.5475-5501.
Kim, W., Goyal, B., Chawla, K., Lee, J., Kwon, K., 2018.
Attention-based ensemble for deep metric learning. In
The European Conf on Computer Vision (ECCV).
Lin, Z., Yang, Z., Huang, F., Chen, J., 2018. Regional
maximum activations of convolutions with attention for
cross-domain beauty and personal care product
retrieval. ACM Multimedia Conf., pp.2073–2077.
Mohedano, E., Salvador, A., McGuinness, K., Marques, F.,
O’Connor, NE., A. Salvador, F., Giro-i Nieto, G., 2016.
Bags of local convolutional Features for Scalable
Instance Search.
Nazir, A., Ashraf, A., Hamdani, T., Ali, N., 2018. Content
based image retrieval system by using hsv color
histogram, discrete wavelet transform and edge
histogram descriptor. In International Conference on
Computing. Mathematics and Engineering
Technologies (iCoMET), pp.1-6.
Nguyen, N., Rigaud. C., Burie. JC., 2018. Digital Comics
Image Indexing Based on Deep Learning. J. Imaging.
vol.4(7), pp.89.
Paulin, M., et al., 2015. Local convolutional features with
unsupervised training for image retrieval. In ICCV.
Piras, L., Giacinto, G., 2017. Information fusion in content
based image retrieval: A comprehensive overview.
Information Fusion. vol.37, pp.50-60.
Qassim, H., Verma, A., Feinzimer, D., 2018. Compressed
Residual-VGG16 CNN Model for Big Data Places
Image Recognition. IEEE 8th Annual Computing and
Communication Workshop and Conference (CCWC).
Qin, J.H., et al.,., 2015. Ceramic tile image retriveval
method based on visual feature. Journal of
Hybrid Shallow Learning and Deep Learning for Feature Extraction and Image Retrieval
171
Computational and Theoretical Nanoscience. vol.12,
pp.4191-4195.
Radenovic, F., Tolias, G., Chum, O., 2017. Fine-tuning
CNN Image Retrieval with No Human Annotation.
IEEE Transactions on Pattern Analysis and Machine
Intelligence.
Razavian, A., Azizpour, H., Sullivan, J., Carlsson, S., 2014.
CNN features off-the-shelf: an astounding baseline for
recognition. Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition Workshops,
pp.806-813.
Razavian, A.S., Sullivan, J., Carlsson, S., Maki, A., 2016.
Visual instance retrieval with deep convolutional
networks. ITE Transactions on Media Technology and
Applications, vol. 4(3), pp.251-258.
Sun, G., Wang, C., Ma, B., Wang, X., 2017. An improved
SIFT algorithm for infringement retrieval. Multimedia
tools and Applications.
Varish, K.N., Pal, A.K., 2015. Content based image
retrieval using statistical features of color histogram. In
International Conference on Signal Processing,
Communication and Networking (ICSCN), pp.1-6.
Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y.,
Manzagol, P.A., 2010. Stacked Denoising
Autoencoders: Learning Useful Representations in a
Deep Network with a Local Denoising Criterion. J.
Mach. Learn. Res. vol. 11, pp.3371-3408.
Wang, J.Z., Li, J., Wiederhold, G., 2001. SIMPLIcity:
Semantics-Sensitive Integrated Matching for Picture
Libraries. IEEE Transactions on Pattern Analysis and
Machine Intelligence. vol.23 (9), pp.947-963.
Wu, L., Wang, Y., Gao, J., Li, X., 2018. Deep adaptive
feature embedding with local sample distributions for
person re-identification. Pattern Recognition. vol.73,
pp.275-288.
Wu, L., Wang, Y., Li, X., Gao, J., 2018. Deep Attention-
Based Spatially Recursive Networks for Fine-Grained
Visual Recognition. IEEE Transactions on Cybernetics.
vol.49, pp.1791-1802.
Wu, Y., Wu, Y., 2009. Shape-Based Image Retrieval Using
Combining Global and Local Shape Features
International Congress on Image and Signal
Processing, Tianjin, pp.1-5.
Yuan, X., Yu, J., Qin, Z., Wan, T., 2011. A sift-lbp image
retrieval model based on bag-of- features. Int.
conference on image processing, pp.1061-1164.
Yu, J., Qin, Z., Wan, T., Zhang, X., 2013. Feature
integration analysis of bag-of-features model for image
retrieval. Eurocomputing, vol. 120, pp. 355-364.
Zhou, W., Li, H., Tian, Q., 2019. Recent Advance in
Content-based Image Retrieval: A Literature Survey.
ICEIS 2020 - 22nd International Conference on Enterprise Information Systems
172
