University of Nebraska - Lincoln

DigitalCommons@University of Nebraska - Lincoln

Library Philosophy and Practice (e-journal) Libraries at University of Nebraska-Lincoln

April 2023

The FASHION Visual Search using Deep Learning Approach

Smita V. Bhoir
Department of Computer Engineering, K.J. Somaiya College of Engineering, Ramrao Adik Institute of
Technology, Mumbai, Maharashtra, India, [email protected]

Sunita R. Patil
2Department of Computer Engineering, K.J. Somaiya Institute of Engineering & Information Technology,
Mumbai, Maharashtra, India, [email protected]

Follow this and additional works at: https://digitalcommons.unl.edu/libphilprac

Part of the E-Commerce Commons, Other Computer Engineering Commons, Scholarly

Communication Commons, and the Scholarly Publishing Commons

Bhoir, Smita V. and Patil, Sunita R., "The FASHION Visual Search using Deep Learning Approach" (2023).
Library Philosophy and Practice (e-journal). 7569.
https://digitalcommons.unl.edu/libphilprac/7569
 The FASHION Visual Search using Deep Learning Approach

Smita Bhoir1, a *, Sunita Patil2,b

1
Department of Computer Engineering, K.J. Somaiya College of Engineering, Mumbai,
Maharashtra, India
2
Department of Computer Engineering, K.J. Somaiya Institute of Engineering & Information
Technology, Mumbai, Maharashtra, India
a
[email protected], [email protected]

Keywords: E-commerce; Visual Search; Deep Fashion Convolution Neural Network; Fashion
Classification;

Abstract. In recent years, the World Wide Web (WWW) has established itself as a popular source of
information. Using an effective approach to investigate the vast amount of information available on
the internet is essential if we are to make the most of the resources available. Visual data cannot be
indexed using text-based indexing algorithms because it is significantly larger and more complex than
text. Content-Based Image Retrieval, as a result, has gained widespread attention among the scientific
community (CBIR). Input into a CBIR system that is dependent on visible features of the user's input
image at a low level is difficult for the user to formulate, especially when the system is reliant on
visible features at a low level because it is difficult for the user to formulate. In addition, the system
does not produce adequate results. To improve task performance, the CBIR system heavily relies on
research into effective feature representations and appropriate similarity measures, both of which are
currently being conducted. In particular, the semantic chasm that exists between low-level pixels in
images and high-level semantics as interpreted by humans has been identified as the root cause of the
issue. There are two potentially difficult issues that the e-commerce industry is currently dealing with,
and the study at hand addresses them. First, handling manual labeling of products as well as second
uploading product photographs to the platform for sale are two issues that merchants must contend
with. Consequently, it does not appear in the search results as a result of misclassifications. Moreover,
customers who don't know the exact keywords but only have a general idea of what they want to buy
may encounter a bottleneck when placing their orders. By allowing buyers to click on a picture of an
object and search for related products without having to type anything in, an image-based search
algorithm has the potential to unlock the full potential of e-commerce and allow it to reach its full
potential. Inspired by the current success of deep learning methods for computer vision applications,
we set out to test a cutting-edge deep learning method known as the Convolutional Neural Network
(CNN) for investigating feature representations and similarity measures. We were motivated to do so
by the current success of deep learning methods for computer vision applications (CV). According to
the experimental results presented in this study, a deep machine learning approach can be used to
address these issues effectively. In this study, a proposed Deep Fashion Convolution Neural Network
(DFCNN) model that takes advantage of transfer learning features is used to classify fashion products
and predict their performance. The experimental results for image-based search reveal improved
performance for the performance parameters that were evaluated.

1. Introduction
The fashion industry is expanding at a rapid pace in both the United States and internationally.
In comparison to 2020, the domestic fashion market is predicted to grow by 8.8 percent in 2024,
reaching USD (United States dollars) $26.288 million in value. [1][2]. The apparel and fashion
online market are expected to reach USD 113 million in 2018, and it is growing at a rate of 18
percent per year, with mobile shopping accounting for approximately 40% of the total.
The fashion industry also includes accessories such as purses, shoes, and jewellery, which are
categorized as such. The automation, development, and digitization of manufacturing and
distribution have resulted in a new paradigm for the fashion industry. As the fashion industry
becomes more digital, the system evaluates customer fashion desires and attempts to match
those desires as quickly as possible. As a result, digital technology in the fashion industry is
attracting the attention of a diverse range of consumers who, as a result of the shorter
manufacturing cycle, have access to a greater variety of products. Customers' use of smart
gadgets has increased to more than 10 million as a result of the introduction of a new e -
commerce platform [3]. Traditional garment sales channels such as brick-and-mortar stores
have seen sales decline, whereas non-store retailers such as online and mobile merchants have
seen sales increase at the fastest rate. When compared to previous years, the online f ashion
market increased by 16 percent per year from 2011 to 2016, and from January to September
2017, the market increased by 20 percent over the previous year [4]. Eventually, it is expected
to become a major channel in the fashion industry in the coming years.
Since it provides a hassle-free shopping experience and delivery to the user, e-commerce has
transformed the world of consumerism and triggered an increase in demand for goods. Because there
is a greater selection of fashion products available on e-commerce platforms than in traditional retail
stores, shopping for fashion apparel on e-commerce platforms differs significantly from shopping in
traditional retail stores. As a result, a system that more efficiently assists consumers in searching
for the goods they desire and recommending the desired product is becoming increasingly
important to businesses. Therefore, if you want to communicate effectively in the fashion
industry, you cannot rely solely on verbal communication because visual cues and desig n
elements are so important. Although the existing fashion retrieval system employs a text -based
retrieval approach that is based on product attribute information (product name, category, brand
name, and so on), the system is not without its flaws (product name, category, brand name,
etc.). A text-based search strategy has limitations when it comes to providing appropriate search
results in the fashion industry, which includes a significant design component. Numerous
"shopping how" and "smart lens" systems have been developed in recent years, but they have
all failed to deliver good or noteworthy results when searching for products using fashion
images as a starting point. Users must either find an image of the goods online or take their
photographs to use these systems.
Two major issues affect the industry, which is discussed from both the seller's and buyer's
perspectives in this study [5]. An e-commerce site requires sellers to upload photos of their
products as well as relevant labels to the product's description for the product to be sold.
Because of the involvement of humans, this process is prone to errors. Product misclassification
can cause products to be missed in search results, resulting in lower or no sales for the company
that makes them. By using machine learning models, the photographs can be classified with
high accuracy, which in turn motivates the vendors to categorize them correctly. In addition, a
customer's demand may be delayed because he does not understand the correct terminology [5].
When a customer is shopping on an e-commerce website, he or she typically enters the product's
keywords into the search bar. Its search algorithm compares the keywords entered by the user
with the product labels stored in its database, and it returns results that are relevant to the user's
inquiry. When the user locates the product he or she is looking for, the user orders it directly
from the search results page. Using a text-based search necessitates that the consumer has a
thorough understanding of the product and is aware of the terms that should be entered into the
search toolbar. It's important to note that this isn't always the case. We come into contact with
a wide variety of things in our daily lives that we are completely unaware of. Occasionally, w e
are unable to conduct a product search on the e-commerce website due to technical difficulties.
By employing visual search techniques, you can overcome this difficulty. Shoppers who
conduct visual searches look for products that include images or other v isual cues rather than
products that contain keywords. Customers can search for similar products by simply taking a
picture of what they want and uploading it to the visual search engine.
Any visual search algorithm can benefit from machine learning models, which can be used to learn
attributes about new photographs and search for comparable products. To provide more relevant
search results, additional images with similar features will be added to the visual search engine once
the target image has been uploaded to make the search results more relevant as well. It is possible to
generate images' latent properties using autoencoders and other visual search methods, for example.
To further enhance the retrieval of image embedding features, deep neural network models that
have already been trained can be applied to the problem [4][6][7].
A large number of images from all over the Internet are available to users through social media
and respectable smartphones, which means that users have access to a large number of images
from all over the Internet at once. In these circumstances, the ability to search for, filter, and
organize photos becomes increasingly important. Manually searching for the images you want
is an option if you have a small number of images. As the number of things grows, this becomes
impossible to accomplish. To keep up with this rapid expansion, it is necessary to develop
picture retrieval systems that are capable of operating across a wide range of platforms. It is
our goal to develop a retrieval system that is accurate in handling and querying a database of
photographs [8][9].
To summarize, this paper will pursue three broad objectives:
Image Labeling: In E-commerce, while purchasing products appropriate labeling of products is
important. Many images over the web are unlabeled [10].
Image Classification: To design and train different neural network models to learn from large
sets of images of products from an e-commerce website [11].
Image Search: To use autoencoders and cosine similarity to identify similar images [12][13].

2. Transfer Learning
Deep learning [14] is a machine learning method that belongs to a large family of machine
learning methods. [15] Deep learning classifiers, as opposed to traditional neural network
classifiers, develop classifiers with numerous hidden layers to identify the salient low -level
features of a picture, as opposed traditional neural network classifiers. Concerning Deep
Learning, a technique known as transfer learning makes use of an artificial neural network's
ability to use features learned from a previous problems to solve a new problem within the same
domain. It is advantageous to learn through transfer[16][17][18] for a variety of reasons.
• First, it saves computational time by reusing information from a previous training
process rather than starting from scratch with a new model, rather than starting from
scratch with a new model [16].
• The second advantage is that it builds on the knowledge and experience gained through
the use of previous models [17].
• The third point to mention is that when the new training dataset is small, transfer learning
is extremely beneficial [18].
Transfer learning has the potential to benefit a wide range of applications, including computer
vision, audio categorization, and natural language processing. Many attempts have been made to
automate the classification of images, either to speed up the process or to improve accuracy. To
solve the picture categorization problem, the convolutional neural network (CNN) was
developed as one of the first approaches. Because of the work of Krizhevsky et al.[19], CNN
was able to outperform all other methods for solving the picture classification problem by the
year 2012. They achieved state-of-the-art performance in the ImageNet Large Scale Visual
Recognition Challenge (ILSVRC) competition, outperforming other commonly used machine
learning algorithms and achieving superior results. In all of these scenarios, training the weights
of the deep network from scratch requires a significant amount of time and a large amount of
data (hundreds of thousands of images). The presence of these constraints makes deep learning
algorithms extremely difficult to implement in the domain of image data, where there are
frequently only a few images available [20]. A significant amount of time and knowledge is
required to annotate fashion images; this is where transfer learning may be beneficial to
students. Utilizing an architecture that has already been trained eliminates the need to start from
scratch. According to several studies, CNN has been used to identify fashion images, either
through transfer learning or the introduction of novel architectural structures. According to [24],
deep CNN-based fashion image categorization models are presented in [21–35] of the literature
reviewed. Some critical open questions are presented here to improve the utilization of transfer
learning in the fashion industry.

3. CNN Architectures used in Transfer Learning

The main CNN architectures used in transfer learning are discussed in detail in this section.
According to [36], the rise of deep learning in image classification began in 2012 with the
introduction of AlexNet [14], which included the ReLU activation layer as well as other deep
learning techniques. It was discovered that using a CNN to classify images improved its
accuracy while also eliminating the need to manually feature engineer each image. Following
the introduction of AlexNet, a slew of new architectures, including VGG16, VGG19, ResNet,
GoogLeNet, DenseNet, and Xception, were introduced, each with additional features to more
effectively classify images.

A. VGG Network Architecture

The Visual Geometry Group at Oxford University introduced two new structures in
2014, which were designated VGG16 [37] and VGG19 [37]. The VGG16 architecture is
comprised of five convolution blocks and three thick layers, with a total of 138,355,752
parameters in the architecture. Each block contains many convolutional layers, followed by a
max pool layer, which helps to reduce the block output size and remove noise. In total, six
convolutional layers are used in the final three blocks, with two being used in each of the first
and second blocks. One stride in the kernel is used by this network, and it is stride 1. Last but
not least, an additional layer was created that would transform each block into a single -
dimensional object that could then be inserted into the various levels that were connected. Total
connectivity is achieved by connecting 4096 neurons in the first two layers and 1000 neurons
in the third and final layer in the third and final layer. In addition to the fully linked layers, a
softmax layer is added to ensure that an output probability summation is a posi tive number. To
emphasize the difference between the two, VGG19 has 19 convolution layers, whereas VGG16
only has 16. As a result of the addition of layers, the number of parameters increases from
138,357,544 to 143,667,240. In the authors' opinion, the addition of additional layers increases
the system's robustness and ability to learn complex architectures.
The sequential blocks of this network's sequential convolutional layers allow for a
significant reduction in the amount of spatial information that must be stored in the network.
The majority of the time, this network suffers from a lack of feature extraction weights, which
is a significant flaw in the overall design. A large number of additional variables must be taken
into consideration.
B. ResNet Network Architecture
ResNet, which is an abbreviation for residual network, was invented by He et al. [38] in
2015. There are a total of 25,000,000 variables in use. When comparing ResNet to other
architectures, you'll notice that it has a single connection between the convolutional layers and
the ReLU activation layer, which is referred to as a "residual connection." This connection is
what makes ResNet unique among the rest. It is ensured by the residual connection that weights
learned from preceding layers do not disappear during the backpropagation of the network. The
three iterations of this network are designated as ResNet50, ResNet101, and ResNet152 (each
with a different number of layers). The use of residual connections in this network has several
advantages, not the least of which is that it allows for the use of multiple levels. Furthermore,
increasing the depth of the network rather than the width of the network makes it easier to
reduce the number of unnecessary parameters. The fact that the summation that occurs in each
residual block keeps the filter size constant is one of the main disadvantages. Furthermore, to
successfully train this network, large datasets are required, resulting in a time -consuming
training phase that is expensive in terms of computing resources.
C. GoogLeNet Network Architecture
According to Google researchers, the GoogLeNet network, also known as the
IncpetionV1 architecture, was first implemented in 2014 [39]. Following the success of
InceptionV1, the writers went on to create sequels, including InceptionV2 and InceptionV3. To
capture different aspects of images, GoogleLeNet's fundamental idea is to use multiple
convolution layers in the same block to go deeper and wider; these blocks are referred to as
Inception blocks in the GoogLeNet architecture. In the GoogLeNet designs, InceptionV1 and
InceptionV3 are the two most frequently encountered. A total of six convolution layers are used
in the InceptionV1 inception blocks, while a total of seven convolution layers are used in the
InceptionV3 inception blocks (see figure). This network's inception module enables the
convolution layers to be used in parallel, allowing it to collect images with different aspect
ratios at the same time as it collects images with the same aspect ratio. The most significant
weakness of this network is the amount of computing effort required to train it, which is a result
of the network's extensive layer structure.
D. AlexNet Network Architecture
As a starting point, CNN selected AlexNet [40] as the architecture for its participation
in the 2012 ImageNet competition. The AlexNet architecture is made u p of five convolution
layers and three dense layers, for a total of nine layers. In AlexNet, two novel features were
introduced: reLU activation (instead of sigmoid activation) and dropout, both of which were
intended to combat the overfitting that can occur as a result of the deep architecture's deep
learning. The main advantage of this network is that it is computationally efficient when
compared to other networks, which is particularly important during the training stage. A
common misconception is that because of the depth of the network, fine details in images cannot
be detected.
E. DenseNet Network Architecture
DenseNet architecture [41] [42]refers to convolutional networks that are densely
connected and have a large number of connections. Thick blocks, rather than residual
connections, were recommended as a replacement for ResNet by the authors. Convolution
layers are added to the dense block progressively, but each layer has a direct relationship to the
layers that come after it. All of the information from the layers that came before it is fed into
each subsequent layer of convolution. The DenseNet contains a total of 8,062,504 variables.
The fact that all layers are connected prevents information loss, which is a significant advantage
of this network (especially the deep layers). The following are the most significant
disadvantages: it needs a large amount of data to work with during the training phase if you
want to achieve good results.
F. Xception Network Architecture
 Architect Chollet was inspired by the InceptionV3 architecture when he created the
Xception (which stands for extreme inception)[43][44]. Xception's design makes use of two
key concepts: convolution in-depth and pointwise separability. These concepts are used to
replace the inception module, which was previously present. In this 71-layer network, there are
a total of 22.9 million parameters. Comparing this network to other deep networks, its main
advantage is that it has a deep architecture but only a small number of parameters. This is an
important distinction. The network requires a large amount of data from a variety of sources to
train all of the network's parameters properly.
To better understand the significance of transfer learning, experimental work is carried out to
train models such as the VGG16, VGG19, ResNet, GoogLeNet, DenseNet, and Xception,
among others. Table I presents a summary of the results obtained using the number of
parameters and accuracy of the parameters over the ImageNet dataset. The accuracy of a
classification system is determined by dividing the number of correctly classified observations
by the total number of observations. Accuracy over predicted labels y is measured as top -k
accuracy, with the top 5 accuracies representing accuracy over five classes, and the top 1
accuracy representing accuracy for a single-class classification.

Table I: Top 5 accuracy, top 1 accuracy, and the number of parameters of AlexNet, VGG,
Inception, and ResNet in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC)
challenge.
Architecture Number of Top-5 Top-10
Parameters Accuracy Accuracy
Xception 22,910,480 84.60% 63.30%
VGG-19 138,357,544 91.90% 74.40%
GoogLeNet 23,000,000 92.2% 74.80%
ResNet‐152 25,000,000 94.29% 78.57%
DenseNet 8,062,504 93.34% 76.39%
AlexNet 62,378,344 94.50% 79.00%

4. Proposed Work and Methodology

The research proposal uses structure, the framework to work on the technology. Our approach
incorporates three modules. Background removal and extraction of valuable features from the
image are the first two modules in the process of image processing. After that, the deep CNN
model is trained to examine field-specific feature representations. Finally, the third module uses
the features acquired from the trained model to find images that are comparable to the query
image. The algorithm depicts the general workflow of the project.
 Authentic fashion images from the most popular Indian e-commerce web application, myntra.com,
are used in this work. Because the Fashion Product Images (Small) dataset is readily available on
Kaggle, there is no need to scrape images from the website to complete the task. It also has the
advantage of having color images, which makes it more similar to a real-world business problem than
other types of information. Google Collaboratory, which imports Kaggle data seamlessly, provides
free GPU/TPU resources for convolutional neural network (CNN) algorithms for image
classification. These algorithms can be used to classify images. Python modules such as TensorFlow
and others are already installed, allowing for quick and simple CNN image modeling in Python.
During the cataloging process, the dataset includes a variety of label attributes that were manually
entered to describe each object, as well as high-quality images taken of the objects themselves.
Because the low-resolution images present several computational and run-time issues, we chose the
smaller image collection, which contains the same set of images but at a lower resolution, as an
alternative.
Step 1: Import Libraries
Import the necessary libraries to process the data like (NumPy, pandas, os, OpenCV,
matplotlib, etc.)

Step 2: Loading Dataset

There are 44,424 color channels in the CSV file. Styles.csv can be used to map an image's
metadata to its numeric id. Images IDs and metadata IDs were merged and we found 44,419
instances where they match. These photo sets are being considered for use in modeling. Gender,
Master Category, Sub-Category, Article Type, Season, etc. are all completely labeled in the
dataset. We use three schemes for product classification namely - a concatenated Gender and
Master category, Sub Category, and Article Type; as they are clearly labeled without any
missing values. Load the dataset from the given directory path for further processing and dis play
some samples of images from the dataset.
Step 3: Process CSV File
Information about the product is contained in a CSV file (id, gender, masterCategory,
subCategory, articleType, color, and so on). As demonstrated in Figure 1, Figure 2, Figure 3,
and Figure 4, we read the CSV file and count the number of items based on several categories
such (masterCategory, subCategory, articleType) and produce the bar graph.
Step 4:Pre-process the data
a) NumPy array Conversion
 Images are converted into NumPy Array in Height, Width, Channel format.
 The main advantage of NumPy array Conversion, e.g., You can load the images into a
NumPy array so that you can save computation time. Images can be easily loaded.
b) Generate pixelate form
 Convert NumPy array back to Image format in pixelated form.
c) Image Resize
 Image resizing refers to image scaling. This tends to minimize no. of pixels from the
image. For example, that has other effects; for example, it can minimize the training
time of neural network as more is no. of pixels in the image more is no. of input nodes
that improve the model’s complexity.
d) Data Filtration
 Limitations- Existing research papers surveyed considered categories of products having
maximum images, neglected categories with less number of products [45]
 To give scope for searching newly available products in a category with less number of
products proposed model does not filter any class or images with less number of images.
 Data Filtering is one of the most frequent data manipulation operations. We have used
to filter the images based on some conditions (ratio of the images, images that are not
present in the dataset), during image pre-processing, images are automatically rejected
[46].
e) Gray Scaling
 Gray scaling is the conversion process of an image from other color spaces, for example,
RGB, CMYK, HSV, etc. Dimension reduction and Reduction in model complexity are
benefits of gray scaling [47][48].
gen masterCa subCat article baseC seas yea usa productDisp
id
der tegory egory Type olour on r ge layName

Turtle Check
159 Topwe Navy 201 Cas
0 Men Apparel Shirts Fall Men Navy
70 ar Blue 1.0 ual
Blue Shirt

Peter
393 Bottom Sum 201 Cas England Men
1 Men Apparel J eans Blue
86 wear mer 2.0 ual Party Blue
J eans

Titan
592 Wo Accessori Watche Watch Wint 201 Cas
2 Silver Women
63 men es s es er 6.0 ual
Silver Watch

Manchester
213 Bottom Track 201 Cas United Men
3 Men Apparel Black Fall
79 wear Pants 1.0 ual Solid Black
Track Pants

537 Topwe Sum 201 Cas Puma Men

4 Men Apparel Tshirts Grey
59 ar mer 2.0 ual Grey T-shirt

Inkf ruit Mens

185 Topwe Sum 201 Cas Chain
5 Men Apparel Tshirts Grey
5 ar mer 1.0 ual Reaction T-
shirt

Fabindia
308 Topwe Sum 201 Eth
6 Men Apparel Shirts Green Men Striped
05 ar mer 2.0 nic
Green Shirt

J ealous 21
269 Wo Topwe Sum 201 Cas
7 Apparel Shirts Purple Women
60 men ar mer 2.0 ual
Purple Shirt

Puma Men
291 Accessori Navy Sum 201 Cas
8 Men Socks Socks Pack of 3
14 es Blue mer 2.0 ual
Socks

300 Accessori Watche Watch Wint 201 Cas Skagen Men

9 Men Black
39 es s es er 6.0 ual Black Watch

Figure 1: Sample Input dataset after loading into the model

 Figure 2: Count the Total no of images by masterCategory

Figure 3: Count the Total no of images by SubCategory

Figure 4: Count the Total no of images by ArticleTypeCategory

f) Auto Labelling
 To allow automatic labeling (assign unique values), the model must be enabled to
understand what is depicted in the picture when information is being processed. It has
to be trained to know which tag should be attached to each data unit based on article
type.
g) Normalization
  It is a method often used in machine learning data preparation. The objective of
normalization is to modify the numerical column values of the dataset to a similar scale
without distorting the values or missing information.
h) Creating a List with Unique Value
 Created a list with unique values and added a column with their specified element
position. Store the Labeled dataset for training the model.
• We create an augmented data frame with a file name and also mapping with product id.
• We also perform one of the important tasks called labeling, we assign the unique id to
each article Type.
• Gray Scaling is the process to Convert the colored images into grayscale images.

Normalization is a technique that is commonly used in data processing for machine learning.
The goal of normalization is to transform the values of the dataset's numeric columns to
comparable scales without distorting ranges of values or losing information. Figure 5 shows
sample output images.

Figure 5: Sample output images

Step 5: Designing Proposed Deep Fashion Convolution Neural Network (DFCNN)

 The proposed layered architecture of Deep Fashion Net is given in Figure 6. The design
has 8 layers: 5 convolutional layers and 3 fully linked layers.

Figure 6: Layered architecture of proposed Deep Fashion Convolution Neural Network

(DFCNN)
• Working details of Deep Fashion Net Model
The proposed DFCNN (Deep Fashion Convolutional Neural Network) is an 8 -layer
architecture that comprises five convolutional layers, and 3 dense layers. Input to this model is
images of size 227X227X3. Figure 6 shows Layer-wise working details of Proposed DFCNN
Model. The first layer of max-pooling, size 2*2, and step 2 are now available in the following.
The resultant feature map in size 27X27X96 is then provided. After this, the second
convolution operation is applied. We have 256 filters of this type in the filter size, 11 * 11. The
step is 1 and ReLu is used again to activate the function. Now the output size we get is
17X17X256. Next, we have a second 2*2 size and step 2 max-pooling layer. Then we have
8X8X256 as the output map. Now we use the third convolution procedure with 384 filters of 3*3
size. The step is one. Again, ReLu is used for the activation function. The output feature map
is in a 6X6X384 shape.The fourth convolution operation is then applied with 384 filters 3*3 in
size. The step is one. Again, ReLu is used for the activation function. The output feature map
is 4X4X384 in shape. We then get the last 3*3 convolution layer with 256 filters of this kind.
The step is 1 and the activation function is ReLu. The output feature map is 2X2X256. If the
architecture is observed so far, there are more and more filters as we go deeper. Hence it is
extracting additional features as we get further into the architecture. Furthermore, the filter size
is lowering and the initial filter is larger and the filter size decreases as we proceed and the
feature map shape is decreasing. Now, we apply the 3rd max-pooling layer of size 2*2 & stride
2. The output of the feature map shape is 1X1X256. Flatten is the function to transform the
pooled map into a single column which is sent to a fully connected layer. Dense will add the
fully connected neural network layer. We have a 1st fully connected layer with a ReLu
activation function. The output dimension is 4096. The first drop-out layer comes next with a
drop-out rate of 0.4 fixed. 2nd fully connected layer with 4096 neurons & ReLu activation
function. Next follows the second drop-out layer with a 0.4 drop-out rate. 3rd fully connected
layer with 1000 neurons & ReLu activation function. The third drop-out layer comes in the next
one with 0.4 drop-out rates. Finally, there is a layer or output layer fully connected with no.
of class in the dataset. This layer uses Softmax for the activation function.
Step 6: Split the dataset
• We split the dataset into 80:20, 70:30, and 50:50 training or testing datasets.
Step 7: Model Training
The DFCNN model is trained and results are obtained for multiclass classification under
SubCategory, MasterCategory, ArticleTypeCategory, also, a combination of categories under
Gender+Master Category.
Step 8: Performance Evaluation
The performance of learning algorithms on test data is typically used to determine the quality
of the algorithms [38]. The first is the Receiver Operating Characteristic Curve (ROC), also
known as the Area Under the Curve (AUC), which is a widely used performance measure in
supervised classification and is based on the relationship between sensitivity and specificity
[39]. The second is the Receiver Operating Characteristic Curve (ROC), also known as the Area
Under the Curve (AUC), which is a widely used performance measure in unsupervised
classification. It was decided to use a generalization of the AUC for multiple classes in this
study. Hand and Till [49][50][40] defined a function that performs multi -class AUC and is
composed of the mean of several AUC values. This function is defined as follows: This can be
calculated using Equation (1), where TP denotes the true positives, which are the number of
instances that are positive and are correctly identified, and FN d enotes the false negatives,
which are the number of positive cases that are incorrectly classified as negative.
Sensitivity (Recall) = TP ( 1)
TP + FN
If the conditional probability of true negatives is given a secondary class, specificity
corresponds to the probability of the negative label being true, which means that it approximates
the probability of the negative label being false; where TN is the number of true positives or
negative cases that are negative and correctly classified as negative, and FP is defined as the
number of false positives, defined as the number of negative instances that are incorrectly
classified as positive cases.
Specificity = TN (2)
TN + FP
In general, sensitivity and specificity are used to evaluate the effectiveness of an algorithm
on a single class of data, with positive and negative values being used to indicate effectiveness.
The accuracy of a classification system is the most commonly used metric to evaluate
classification performance. When it came to the evaluation stage, the accuracy was calculated
after every 20 iterations of the procedure. Calculated as a percentage of correctly classified
samples, this metric can be expressed in the following way:

Accuracy = TP + TN (3)
TP + TN + FP + FN
When it comes to precision, it's defined as the number of true positives divided by th e sum
of true positives and false positives. This measure is concerned with correctness, i.e., it assesses
the algorithm's ability to predict the future. Precision refers to how "precise" the model is in
predicting positive outcomes and how many of those predictions are correct.
 Precision = TP (4)
TP + FP
It is determined by the harmonic mean precision and recall, which is called the F-score. It is
primarily concerned with the analysis of positive classes. This metric has a high value, which
indicates that the model performs better when compared to the negative class.
F-score =2*Precision * Recall (5)
Precision + Recall

5. Experimental Results & Analysis

We plot the accuracy evolution graph, Training and validation accuracy graph, confusion
matrix, and model accuracy graph. We also display 3-4 results on internal images and 5-6 results
on external images based on different classes (shirts, T-shirts, watches, sarees, jeans, jewelry,
etc.). A confusion matrix is a table i.e. commonly applied to define the perfor mance of a
classification model (or "classifier") onset of known test data. The confusion matrix itself is
simple to understand, but the verbiage connected with it may be confusing.
A. Overall Performance Analysis
The performance analysis of Master Category, ArticleType, SubCategory, and MasterGender
Category is given in Table II, Table III, Table IV, and Table V.
Table II. Master Category based Classification
Category Total Splitting Training Validati No. of
dataset Accuracy on epochs
classes
Accuracy
masterCategory 7 80:20 0.99 0.98 50
masterCategory 7 70:30 0.99 0.97 50
masterCategory 7 50:50 0.99 0.96 50
Table III. SubCategory based Classification
Category Total Splitting Training Validation No. of
classes dataset Acc Acc epochs
subCategory 45 80:20 0.99 0.90 50
subCategory 45 70:30 0.99 0.93 50
subCategory 45 50:50 0.99 0.91 50

Table IV. Gender+ Master Category based Classification

Category Total Splitting Training Validation No. of
classes dataset Acc Acc epochs

Gender+masterCategory 23 80:20 0.99 0.88 50

Gender+masterCategory 23 70:30 0.98 0.85 50

Gender+masterCategory 23 50:50 0.98 0.85 50

 Table V. Article-type Category based Classification
Category Total Splitting Training Validation No. of
classes dataset Acc Acc epochs
articleType 143 80:20 0.98 0.81 50
articleType 143 70:30 0.97 0.79 50
articleType 143 50:50 0.98 0.74 50

B. Category wise Detailed Experimental Results

Master Category
The performance analysis of the Master Category is given in Table II.
• Displaying total products

# total no. of product based on masterCategory

df['masterCategory'].value_counts()

Apparel 21397
Accessories 11274
Footwear 9219
Personal Care 2403
Free Items 105
Sporting Goods 25
Home 1

Figure 7: The Graph of Count total no of images by Master Category

 Figure 8: Master Category Loss-Accuracy Evolution Graph

Figure 9: Training and Testing Validation Accuracy and Loss

Figure 10: Master Category Classification Confusion Matrix

• Per class Performance Evaluation
For each class in MasterCategory, the performance is evaluated as given below.
 ID precision recall f1-score support
0 0.98 0.97 0.98 10641
1 0.94 0.95 0.95 5668
2 0.99 0.99 0.99 4610
3 0.86 0.95 0.90 1225
4 0.19 0.11 0.14 56
5 1.00 0.10 0.18 10
Accuracy-0.98

• Prediction Results
The results of the predictions are given in Figure 11. PC indicates predicted class and TC
indicates True Class of product.

Figure 11: Prediction Results

Similarly, results are obtained by the training model in 50:50, 70:30 and 80:20. Figure 12,
Figure 13 and Figure 14 shows Accuracy-Loss Evolution graphs of each categories under
experimental work.

Figure 12: Sub Category Accuracy-Loss Evolution Graphs

Figure 13: Gender+MasterCategory Accuracy-Loss Evolution Graph

 Figure 14: ArticleType Accuracy-Loss Evolution Graph

C. Search Results using DFCNN

The cosine similarity approach is used to evaluate the performance of our visual search
models. A measure of cosine similarity between two non-zero vector arrays in a vector space is
computed by taking the cosine of the angle between these two vectors and computing their
similarity. The result is neatly bounded in the range [0,1], with 1 indicating that two vectors
with the same orientation are similar, and 0 indicating that two vectors with opposite
orientations are dissimilar. The visual search models are also evaluated qualitatively, with the
following steps: 1) random selection of a target image to get recommendations from the current
dataset; 2) comparison of the recommended image labels (gender, masterCategory, sub -
category, and articleType) against the target image; and 3) using external imag es as targets and
retrieving the recommended image labels from the current dataset. During the visual search
process, the top 5, top 10, and top 20 matching images are recommended based on the CNN,
autoencoder, and DFCNN embedding models, respectively. The results of the image search
using both internal and external images are depicted in Figures 16, 17, and 18. The target image
is the image on the left-hand side of each box, and the five images on the right-hand side of
each box are the results of the recommendation using the autoencoder model.
1) Internal Image as Input
• The internal image is given as input to obtain search results.

Figure 15: Auto-labeling of product

• Internal Image as Input with less number of products
The proposed model is designed to obtain results for products with the minimum number
of product images.
Figure 16: Search Result for the product with less number of products in the category using
DFCNN

• Internal Image Input-Top-20 Results

Figure 17: Top-20 search results with DFCNN

2) External Image as Input

 The external image is given as input that is not present in the dataset and search results are
obtained.

op 5 Res lts
xter al ma e as p t

Figure 18: Search results for External image or photograph given as input

5. Conclusion & Future Work

The E-commerce industry necessitates the use of optimized visual search to facilitate the
retrieval or searching of products. The application of machine learning models in enabling
visual search, which solves the problem of searching for products when the correct keywords
are not known by the user, is particularly interesting to researchers. Because it allows users to
search for similar products simply by taking a photo of the product in question, it improves the
customer experience for everyone who uses it. Machine learning models can also be extremely
useful in improving the experience of sellers when they are listing their products on a platform.
Sellers can upload photos of their products, and automated image-to-text machine learning
algorithms can generate appropriate tags to label the products on the marketplace for them. This
can help to reduce inaccuracies in product labeling, which can hurt demand because the products
are not displayed correctly in search results when they are not rendered correctly. The proposed
Deep Fashion Convolution Neural Network (DFCNN) model improved the performance and
accuracy of search results by using deep fashion convolution neural networks. In previous
models, the limitation was found to be that if the number of products was less than a certain
threshold, those products were ignored. With DFCNN model, this limitation is no longer
present. In addition, the computation time is reduced. Indexing method implemented makes
searching faster. When internal and external images are provided as input, along with
appropriate product labeling, the autoencoders developed in conjunction with the DFCNN
model outperform search results in terms of performance.
Machine learning models may also prove useful in the future when it comes to improving the
experience of sellers when they are listing their products on the platform. Sellers can upload
photos of their products, and automated image-to-text machine learning algorithms can generate
appropriate tags to label the products on the marketplace for them. This can help to reduce
inaccuracies in product labeling, which can have a negative impact on demand because the
products are not displayed correctly in search results when they are not rendered correctly. To
accomplish this, CNN models must be combined with natural language processing techniques
such as Word2vec to predict text information from image data and its features. Image
classification may also be used in the identification of counterfeit products, which is another
possible application. An in-depth examination of the characteristics of a brand's logo, such as
its design, colors, placement, and so on, can aid in the identification of bogus products. As
computing power and machine learning algorithms continue to advance at a rapid pace, we may
even be able to use generative adversarial networks (GANs) to come up with new designs for
fashion accessories, thus reducing our reliance on human creativity. Despite the fact that
training a GAN model is extremely difficult, GAN models in the fashion industry have the
potential to provide significant business value very quickly.

Acknowledgments
Ms. Smita Bhoir
M.E. Computer Engineering, PhD (Pursuing). She is pursuing her PhD from K.J. Somaiya College
of Engineering, Mumbai , India. She is having 12 years of R & D experience. Published 15 research
papers, delivered 45 technical talks, organized 40workshops and training programs. Her area of
specialization is Data Science, Deep Web Mining, Wireless Networks.
Dr. Sunita Patil
PhD. Computer Engineeering .She is working as Vice Principal, Dean Academics, Professor,
Department of Computer Engineering, K.J. Somaiya IEIT, Sion, Mumbai, India. She is having more
than 20 years of R & D experience. She has published 30 research papers, organized more than 50
workshops, delivered more than 50 technical talks. Her area of specialization is Data Mining.

REFERENCES

[1] H. Ho a d J. H. Ka , “ he impact of moral philosophy a d moral i te sity o p rchase

behavior toward s stai able textile a d apparel prod cts,” Fash. Text., vol. 6, no. 1, Dec. 2019, doi:
10.1186/s40691-019-0170-8.
[2] S. Y. Kim and J. Ha-Brookshire, “ vol tio of the Korea Marketplace From 1896 to 1938: A
Historical vesti atio of Wester Clothi Stores’ Retail a d Competitio Strate ies,” Cloth. Text.
Res. J., vol. 37, no. 3, pp. 155–170, Jul. 2019, doi: 10.1177/0887302X19835967.
[3]S. Joo and J. Ha, “Fashio d stry System a d Fashio Leaders i the Di ital ra,” J. Korean
Soc. Cloth. Text., vol. 40, no. 3, pp. 506–515, 2016, doi: 10.5850/JKSCT.2016.40.3.506.
[4] J. Jo, S. Lee, C. Lee, D. Lee, a d H. Lim, “Developme t of fashio prod ct retrieval and
recomme datio s model based o deep lear i ,” Electron., vol. 9, no. 3, pp. 1–12, 2020, doi:
10.3390/electronics9030508.
[5] P. Ka r a d R. K. Si h, “A efficie t a alysis of machi e lear i al orithms i CB R,” Proc.
Int. Conf. Comput. Autom. Knowl. Manag. ICCAKM 2020, pp. 18–23, 2020, doi:
10.1109/ICCAKM46823.2020.9051528.
[6] L. Fe zi, S. Ka t, S. Araki, S. Ba era, a d S. S. Sh kla, “Ne ral Networks for Fashio ma e
Classificatio a d Vis al Search,” SSRN Electron. J., 2020, doi: 10.2139/ssrn.3602664.
[7] J. Sa , C. X , a d D. L , “Lear to perso alized ima e search from the photo-shari websites,”
IEEE Trans. Multimed., vol. 14, no. 4 PART1, pp. 963–974, 2012, doi: 10.1109/TMM.2011.2181344.
[8] N. K mar, S. Awasthi, a d D. ya i, “Web Crawler Challe es a d heir Sol tio s,” vol. 7, o.
12, pp. 95–99, 2016.
[9] X. Zha , Z. Li, a d W. Chao, “ mprovi ima e ta s by exploiti web search res lts,”
Multimed. Tools Appl., vol. 62, no. 3, pp. 601–631, 2013, doi: 10.1007/s11042-011-0863-5.
[10] P. S., “A ma e Crawler for Co te t-Based ma e Retrieval System,” Int. J. Res. Eng. Technol.,
vol. 02, no. 11, pp. 33–37, 2013, doi: 10.15623/ijret.2013.0211006.
[11] M. Va ač et al., “Crawli ima es with web browser s pport,” 2015 IEEE 13th Int. Sci. Conf.
Informatics, INFORMATICS 2015 - Proc., pp. 286–289, 2016, doi:
10.1109/Informatics.2015.7377848.
[12] X. Li , J. Li, J. Wa , a d Z. Li , “MMFashio : A Ope -Source Toolbox for Visual Fashion
A alysis,” pp. 1–4, 2020, [Online]. Available: http://arxiv.org/abs/2005.08847.
[13] V. ya i, “Co te t-Based ma e Retrieval ech iq es: A Review,” Content-Based Image Retr.,
pp. 29–48, 2017, doi: 10.1007/978-981-10-6759-4_2.
[14] N. C. Mithun, R. Panda, and A. K. Roy-Chowdh ry, “Co str ctio of Diverse ma e Datasets
from Web Collectio s with Limited Labeli ,” IEEE Trans. Circuits Syst. Video Technol., vol. 30,
no. 4, pp. 1147–1161, 2020, doi: 10.1109/TCSVT.2019.2898899.
[15] W. Zho , H. Li, a d Q. ia , “Rece t Adva ce i Co te t-based Image Retrieval: A Literature
S rvey,” pp. 1–22, 2017, [Online]. Available: http://arxiv.org/abs/1706.06064.
[16] A. . A. M. Alomairi a d G. S lo , “A overview of co te t-based image retrieval tech iq es,”
J. Theor. Appl. Inf. Technol., vol. 84, no. 2, pp. 215–223, 2016.
[17]A. Krizhevsky, . S tskever, a d G. . Hi to , “ ma eNet Classificatio with Deep
Co vol tio al Ne ral Networks.” Accessed: J . 03, 2021. [O li e]. Available:
http://code.google.com/p/cuda-convnet/.
[18] W. L , L. Li, J. Li, . Li, H. Zha , a d J. G o, “A m ltimedia i formatio f sio framework
for web ima e cate orizatio ,” Multimed. Tools Appl., vol. 70, no. 3, pp. 1453–1486, 2014, doi:
10.1007/s11042-012-1165-2.
[19] R. Xia, Y. Pa , H. Lai, C. Li , a d S. Ya , “S pervised Hashi for ma e Retrieval via ma e
Represe tatio Lear i .” Accessed: J . 03, 2021. [O li e]. Available: www.aaai.or .
[20] K. Lin, H.-F. Yang, J.-H. Hsiao, and C.-S. Che , “Deep Lear i of Bi ary Hash Codes for
Fast ma e Retrieval.” Accessed: J . 03, 2021. [O li e]. Available: https://www.cv-
foundation.org/openaccess/content_cvpr_workshops_2015/W03/html/Lin_Deep_Learning_of_2015
_CVPR_paper.html.
[21] C. Szegedy et al., “Goi Deeper with Co vol tio s.” Accessed: J . 04, 2021. [O li e].
Available: https://www.cv-
foundation.org/openaccess/content_cvpr_2015/html/Szegedy_Going_Deeper_With_2015_CVPR_p
aper.html.
[22]J.-C. Chen and C.-F. Li , “Vis al-based Deep Learning for Clothing from Large Database,”
dl.acm.org, vol. 07-09-October, Oct. 2015, doi: 10.1145/2818869.2818902.
[23] N. Khosla a d V. Ve katarama , “B ildi ma e-Based Shoe Search Using Convolutional
Ne ral Networks.” Accessed: J . 04, 2021. [O li e].
Available:http://vision.stanford.edu/teaching/cs231n/reports/2015/pdfs/nealk_final_report.pdf.
[24] Bhoir, Smita V., and Sunita Patil. "A Review on Recent Advances in Content-Based Image
Retrieval used in Image Search Engine." Library Philosophy and Practice (2021): 1-45.
[25] X. Xie, W. Hua , H. H. Wa , a d Z. Li , “ ma e retrieval with adaptive SVM a d ra dom
decisio tree,” Proc. Int. Conf. I-SMAC (IoT Soc. Mobile, Anal. Cloud), I-SMAC 2018, no. 1, pp.
784–787, 2019, doi: 10.1109/I-SMAC.2018.8653699.
[26] N. Kondylidis, M. Tzelepi, and A. exas, “ xploiti tf-idf in deep Convolutional Neural
Networks for Content-Based ma e Retrieval,” Multimed. Tools Appl., vol. 77, no. 23, pp. 30729–
30748, 2018, doi: 10.1007/s11042-018-6212-1.
[27] X. Shi et al., “Pairwise based deep ra ki hashi for histopathology image classification and
retrieval,” Elsevier, Accessed: Jun. 04, 2021. [Online].
Available:https://www.sciencedirect.com/science/article/pii/S0031320318301055?casa_token=Fp3
WRjLtl9MAAAAA:-gdc73HVwoDSM-
kofFFyRcifUb5v3kgc86SppFdoeZzfOTWWKya5lbgqyaLEJrs_mqrY1ZrApk0QLw.
[28] Z. Li, J. a , a d . Mei, “Deep Collaborative mbeddi for Social ma e U dersta di ,”
IEEE Trans. Pattern Anal. Mach. Intell., vol. 41, no. 9, pp. 2070–2083, 2018, doi:
10.1109/TPAMI.2018.2852750.
[29] R. D. Meltser, S. Ba erji, a d A. Si ha, “What’s that Style? A CNN-based Approach for
Classificatio a d Retrieval of B ildi ma es,” 2017 9th Int. Conf. Adv. Pattern Recognition,
ICAPR 2017, pp. 9–14, 2021, doi: 10.1109/ICAPR.2017.8593206.
[30] A. Shah, R. Naseem, Sadia, S. qbal, a d M. A. Shah, “ mprovi CB R acc racy si
co vol tio al e ral etwork for feat re extractio ,” Proc. - 2017 13th Int. Conf. Emerg. Technol.
ICET2017, vol. 2018-Janua, pp. 1–5, 2018, doi: 10.1109/ICET.2017.8281730.
[31] S. Jain and J. Dhar, “ ma e-based search e i e si deep lear i ,” 2017 10th Int. Conf.
Contemp. Comput. IC3 2017, vol. 2018-Janua, no. August, pp. 1–7, 2018, doi:
10.1109/IC3.2017.8284301.
[32] M. R. Kapadia a d C. N. Pa wala, “ mproved CB R System Usi M ltilayer CNN,” Proc.
Int. Conf. Inven. Res. Comput. Appl. ICIRCA 2018, no. Icirca, pp. 840–845, 2018, doi:
10.1109/ICIRCA.2018.8597199.
[33] . A jali, N. Rakesh, a d K. M. P. Akshay, “A Novel Based Decisio ree for Co te t-Based
Image Retrieval: An Optimal Classificatio Approach,” Proc. 2018 IEEE Int. Conf. Commun. Signal
Process. ICCSP 2018, pp. 698–704, 2018, doi: 10.1109/ICCSP.2018.8524326.
[34] F. Sabahi, M. O. Ahmad, a d M. N. S. Swamy, “Percept al ma e Hashi Usi Ra dom
Forest for Content-based Image Retrieval,” 2018 16th IEEE Int. New Circuits Syst. Conf. NEWCAS
2018, pp. 348–351, 2018, doi: 10.1109/NEWCAS.2018.8585506.
[35] R. F , B. Li, Y. Gao, a d W. Pi , “Co te t-based ima e retrieval based o CNN a d SVM,”
2016 2nd IEEE Int. Conf. Comput. Commun. ICCC 2016 - Proc., pp. 638–642, 2017, doi:
10.1109/CompComm.2016.7924779.
[36] N. Bhosle a d M. Kokare, “Ra dom forest-based long-term learning for content-based image
retrieval,” 2016 Int. Conf. Signal Inf. Process. IConSIP 2016, 2017, doi:
10.1109/ICONSIP.2016.7857468.
[37] A. Latif et al., “Co te t-based ima e retrieval a d feat re extractio : A comprehe sive review,”
Math. Probl. Eng., vol. 2019, 2019, doi: 10.1155/2019/9658350.
[38] M. M. Rahma , S. K. A ta i, a d G. R. homa, “A lear i -based similarity fusion and filtering
approach for biomedical ima e retrieval si SVM classificatio a d releva ce feedback,” IEEE
Trans. Inf. Technol. Biomed., vol. 15, no. 4, pp. 640–646, 2011, doi: 10.1109/TITB.2011.2151258.
[39] D. Zha , M. M. slam, a d G. L , “A review o a tomatic ima e a otatio tech iq es,”
Pattern Recognit., vol. 45, no. 1, pp. 346–362, 2012, doi: 10.1016/j.patcog.2011.05.013.
[40] Z. Ze , S. Cai, a d S. Li , “A ovel ima e represe tatio a d lear i method si SVM for
region-based image retrieval,” Proc. 2010 5th IEEE Conf. Ind. Electron. Appl. ICIEA 2010, pp. 1622–
1626, 2010, doi: 10.1109/ICIEA.2010.5514756.
[41] W. S. Wa a d J. J. W , “ rai i asymmetry SVM i ima e retrieval si labeled data,”
Proc. 2009 2nd Int. Congr. Image Signal Process. CISP’09, 2009, doi: 10.1109/CISP.2009.5303323.
[42] J. Donahue et al., “DeCAF: A Deep Co vol tio al Activatio Feat re for Ge eric Vis al
Reco itio .” Accessed: J . 04, 2021. [O li e]. Available: https:// ith b.com/.
[43] D. Ha , Q. Li , a d W. Fa , “A ew ima e classificatio method si CNN tra sfer lear i
a d web data a me tatio ,” Expert Syst. Appl., vol. 95, pp. 43–56, 2018, doi:
10.1016/j.eswa.2017.11.028.
[44] W. Zho , H. Li, Y. L , a d Q. ia , “Lar e scale ima e search with eometric codi ,” i
ACM international Conference on multimedia, 2011, pp. 1349–1352.
[45] J. Philbi , O. Ch m, M. sard, J. Sivic, a d A. Zisserma , “Lost i q a tizatio : mproving
partic lar object retrieval i lar e scale ima e databases,” i IEEE Conference on Computer Vision
and Pattern Recognition, 2008, pp. 1–8.
[46] O. Ch m, J. Philbi , a d A. Zisserma , “Near d plicate ima e detectio : mi -hash and tf-idf
weighting,” i British Machine Vision Conference, vol. 3, 2008, p. 4
[47] Z. W , Q. Ke, M. sard, a d J. S , “B dli feat res for lar e scale partial-duplicate web
ima e search,” i IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 25–32.
[48] W. Zho , Y. L , H. Li, Y. So , a d Q. ia , “Spatial codi for lar e scale partial-duplicate
web ima e search,” i ACM International Conference on Multimedia, 2010, pp. 511–520.
[49] O. Ch m, A. Mik lik, M. Perdoch, a d J. Matas, “ otal recall : Q ery expa sio revisited,”
in IEEE Conference on Computer Vision and Pattern Recognition, 2011, pp. 889–896.
[50] Y. Zha , Z. Jia, a d . Che , “ ma e retrieval with eometry preservi vis al phrases,” i
IEEE Conference on Computer Vision and Pattern Recognition, 2011, pp. 809–816.