See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/360087895

Fashion recommendation system using CNN

Conference Paper · April 2022

CITATION READS

1 4,490

1 author:

Venugeetha Y.
Don Bosco Institute of Technology, Bengaluru, India
15 PUBLICATIONS 6 CITATIONS

SEE PROFILE

All content following this page was uploaded by Venugeetha Y. on 21 April 2022.

The user has requested enhancement of the downloaded file.

 Anjan M. et al.; International Journal of Advance Research, Ideas and Innovations in Technology

ISSN: 2454-132X
Impact factor: 6.078
(Volume 6, Issue 3)
Available online at: www.ijariit.com
Fashion recommendation system using CNN
Anjan M. Abhishek V. C. Balamanikantan
[email protected] [email protected] [email protected]
Don Bosco Institute of Technology, Don Bosco Institute of Technology, Don Bosco Institute of Technology,
Bengaluru, Karnataka Bengaluru, Karnataka Bengaluru, Karnataka
Dheeraj Dr. Venugeetha Y.
[email protected] [email protected]
Don Bosco Institute of Technology, Don Bosco Institute of Technology,
Bengaluru, Karnataka Bengaluru, Karnataka

ABSTRACT recommender systems have taken more and more place in our
lives. From e-commerce (suggest to buyers articles that could
Recommendation systems are the techniques that are used to interest them) to online advertisement (suggest to users the
predict the rating one individual will give to an item or social right contents, matching their preferences), recommender
entity. The items can include books, movies, restaurants and systems are today unavoidable in our daily online journeys.
things on which individuals have different preferences. These One common method when designing systems is content-based
preferences are being predicted using two approaches first filtering.
content-based approach which involves characteristics of an
item and second collaborative filtering approaches which Content-based filtering methods depend on the item
considers user's past behaviour to evaluate its choices. This description and user preference profile. These methods are
thesis proposes a fashion recommendation system which will better suited to situations where there is known data in an
recommend clothing images supported the style sort of the object (name, location, description, etc.), but not the user.
provided clothing images. In this work, we focus on the Content-based treat recommendation as a specific user problem
images of upper body as well as the lower body clothing and and learns the likes and dislikes of a person based on item
with human model in the images. We have created our own features. The algorithms try to recommend items that are
datasets through web scrapping of different e-commerce similar to the user's preferences in the past, or are currently
websites. In this paper we have come up with an idea to build evaluating. It does not rely on a user login machine to generate
a content-based recommendation system using ResNet-50 this temporary profile. In particular, various candidate items
convolutional neural network. are compared to user-rated items and are highly recommended.
This approach has its roots in information retrieval and data
Keywords⸻ Recommendation System, Web scraping,
filtering research.
Vue.js, Flask, ResNet-50, Web App, Content-based filtering
1. INTRODUCTION The aim of the project is to build a model capable of doing
A recommendation program is a comprehensive suite of web fashion recommendation by just looking at its image. The
applications that includes predicting user responses to options. model accepts a image and first determines whether the image
The recommendation system has been a hot topic for a long contains a fashion product or not and recommend it
time. A recommendation program commonly called accordingly. The main objective of this work is to:
Recommendation Techniques, they are simple algorithms that • Develop a fashion recommendation system which answers
aim to provide the most relevant and accurate information to the queries related to fashion shopping.
the user by filtering useful items from a large pool of • To identify the fashion type of given input image.
information resources. Recommendation engines discover data • If the given fashion image is valid then similar set of
patterns in the data set by studying consumer preferences and clothing will be recommended.
generating results related to their needs and interests. In real- • Retrieving the similar search query products from different
time examples like Amazon, they have been using a e-commerce websites.
recommendation engine to suggest products that customers
may like. 2. RELATED WORK
There are some previous works related to building of
During the last few decades, with the rise of YouTube, recommendation systems.
Amazon, Netflix and many other such web services,
© 2020, www.IJARIIT.com All Rights Reserved Page |780
 Anjan M. et al.; International Journal of Advance Research, Ideas and Innovations in Technology
Smart Clothing Recommendation System with Deep Learning The steps to pre-process the image are as follows:
[1] In order to recommend a cloth, we develop two inceptions • Read image: Image provided by the user is taken as the
based convolutional neural networks as prediction part and one input and stored in a temporary folder on the server
feed forward neural network as recommender. In this study, we • Resize image: The saved image is resized in accordance
reach to 98% accuracy on colour prediction, 86% accuracy on with the input size the model is trained with i.e. (224 x 224)
gender and cloth’s pattern predictions and 75% accuracy on • Segmentation: In this stage the saved image is converted
clothing recommendation. from RGB to BGV to aid in better extraction of features.
• Flatten: In this stage after pre-processing the saved, the 2D
Deep Fashion Recommendation System with Style Feature matrix of the image is converted into Vector.
Decomposition [2] Due to the mixed information of style and
category, however, the clothes vector often recommends 3.2.2 Recommendation Engine: A recommendation engine
clothes that do not match. To solve this problem, we propose a filters the information using different algorithms and
style feature extraction (SFE) layer, which effectively recommends the relevant items to users. It first captures the
decomposes the clothes vector into style and category. Based past behaviour of a customer and recommends products which
on the characteristics the category information has small the users might be likely to buy. The working of
variations in the same class while being distinguished from recommendation engine is as follows:
other classes, we extract and remove the category information • Collection of Data: Gathering data is the first step in
from the clothes vector to obtain more accurate style creating the recommendation engine. Data can be either
information. explicit or implicit data. Explicit data can be the input by
users such as ratings and comments on products. And order
3. DESIGN history/return history, Cart events, Page views would be the
3.1 System Architecture implicit data, Click thru and search log. For every user
The proposed system is divided into three parts: visiting the site, dataset will be created.
• Image pre-processing • Analyzing the Data: The filtering of data is done by Real-
• Recommendation Engine time system analysis. The Real-time systems can process data
• Web Scraping as it's created. This system usually involves tools that can
• Web App process and analyze streams of events. It is required to give
These are the four main components. in-the-moment recommendations.
• Filtering the Data: The next step is to filter the data to get
the necessary data to provide recommendations to the user.
Content based filtering approach is used in this project.
Content-based filtering uses meta data or characteristics of
items to recommend other items similar to what the user
likes, based on their previous actions or explicit feedback.
• Web Scraping: Web scraping is the process of automating
the process of data extraction in a fast and efficient manner.
It implements the use of crawlers or robots that automatically
scan specific pages on a website and extract the required
information. For the extraction of product data on a large
scale, we implement a piece of code (called a ‘web scraper’)
that requests a particular product page on an e-commerce
website. In return, the website replies with the requested web
page. Once the page is received, the scraper will parse its
HTML code and extract relevant data from it. When the data
extraction process is completed, the tool finally converts the
data into the desired format. Now, since the web scraper is an
automated program, it can repeat this process thousands of
times on a large number of product pages, and across several
e-commerce websites.
3.2.3 Web App
Front End Design: Vue.js framework is used to create a
interactive interface for the web app.
Back End Design: Flask framework is used to create a
RESTful API.
Fig. 1: System Architecture
4. IMPLEMENTATION
3.2 Proposed System 4.1 Image Preprocessing
3.2.1 Image pre-processing: Image processing is the process The aim of image processing is to enhance the quality of image
of using computer algorithms to perform manipulations on and later on to perform features extraction and classification. It
digital images. The main goal of image processing is to refine is most commonly used in computer vision, medical imaging,
the image data by removing the distorted noise and by meteorology, astronomy, remote sensing and another related
enhancing image pixels. An image is nothing more than a two- field. Tools used in image processing:
dimensional array of number between 0 and 255. It is defined • OpenCV: Using OpenCV we can read and write images,
by the mathematical function f(x,y) where x and y are the two capture and save videos, process images (filter, transform).
co-ordinates horizontally and vertically. The value of f(x,y) at • Numpy: NumPy, which stands for Numerical Python, is a
any point is giving the pixel value at that point of an image. library which consists of multidimensional array objects
© 2020, www.IJARIIT.com All Rights Reserved Page |781
 Anjan M. et al.; International Journal of Advance Research, Ideas and Innovations in Technology
and a numerous tool for processing those arrays. Numpy • Drive Traffic
can be used in processing powerful N-dimensional array • Deliver Relevant Content
object, sophisticated (broadcasting) functions, useful linear • Engage Shoppers
algebra, Fourier transform, and random number
capabilities. 4.3 Web Scraping
• Pillow: Pillow extends the PIL base code to create an Web scraping is used to extract and process large amounts of
improved version which is modern and friendly version of data from the web. The ability to scrape data from the web is a
PIL. It provides tools for opening, manipulating, and saving helpful in procuring huge amounts of data from the web in a
images in various formats. short span of time. Web scraping using Python is an essential
tool you can use to extract the data into a useful form that can
4.2 Recommendation Engine be used for data analysis.
Recommendation Engine may be treated as a black box which
analyze some set of users and shows the items which a single 4.4 Web App
user may like. The major benefits of using a recommendation Web App is used to display information to respective users
engine are obtained from database
• Revenue
• Customer satisfaction

Fig. 2: Steps in Image pre-processing

Fig. 3: Recommendation Engine

Fig. 4: Web Scraping workflow

© 2020, www.IJARIIT.com All Rights Reserved Page |782
 Anjan M. et al.; International Journal of Advance Research, Ideas and Innovations in Technology

Fig. 5: Web App workflow

5. RESULT Using specialized algorithms, product recommendations

This Web App allows the users to find products through an engines are now able to support even the largest of product
image. It provides an easy to use interface which assists the catalogs. The engine is able to intelligently select which
user in finding a product which they might find visually algorithms and filters to apply in any given situation, for any
appealing to them by simply clicking an image of the product given individual shopper. This means that the marketers can
and uploading it on the Web App. The user will be presented maximize conversions and average order value.
with products which the model deemed to be similar to the
uploaded image. The users can buy the product from the 7. REFRENCES
respective E-commerce websites if they wish to do so. [1] Batuhan Aşiroglu, Mehmet Ilkay Atalay, Alkan Balkaya,
Erden Tu Zu Nkan, Mustafa Da Gtekin, Tolga Ensar :
6. CONCLUSION Smart Clothing Recommendation System with Deep
Product recommendations engines are the best way to deliver Learning, 2019 3rd International Symposium on
customers with an improved user experience. Through machine Multidisciplinary Studies and Innovative Technologies
learning, manual curation, and specific algorithms, a product (ISMSIT)
recommendations engine can help bring customers the relevant [2] Yong-Goo Shin, Yoon-Jae Yeo, Min-Cheol Sagong, Seo-
products they want or need. Won Ji, and Sung-Jea Ko: Deep Fashion Recommendation
System with Style Feature Decomposition, 2019 IEEE 9th
It allows marketers to provide customers with relevant product International Conference on Consumer Electronics (ICCE-
recommendations in real time. As a part of an ecommerce Berlin)
personalization strategy, product recommendations [3] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun:
dynamically populate products onto websites, apps, call Identity Mapping in Deep Residual Networks, 2016
centers, or emails, enhancing the customer experience. Cornell University.

© 2020, www.IJARIIT.com All Rights Reserved Page |783

View publication stats