Fashion Recomandation System Using ResNet50

Sadhin, Md. Shanjidul Islam

Dept.CSE MD. ABDUL AWAL MD. NAIM UDDIN JOVY
American International University- Dept.CSE Dept.CSE
Bangladesh American International University- American International University-
Dhaka, Bangladesh Bangladesh Bangladesh
[email protected] Dhaka, Bangladesh Dhaka, Bangladesh
[email protected] [email protected]

Abstract— As people's standard of living went up, they get the picture for this product from images/42431.jpg using
started paying more attention to fashion, which is now seen as a this format. In styles.csv, we've exposed some of the most
popular form of art. Also, the COVID PENDAMIC has made important product categories and their display names for a
people more comfortable with buying things online. Humans are simple start.
inexorably attracted to things that are more aesthetically
appealing. The fashion industry has changed over time because III. RELETED WORK
of the way people like to dress. The fashion industry will be
worth $1,9 billion by the end of 2022. The evolution of the
Recommender systems can recommend products in two
fashion business has resulted from this human nature and an main ways: collaborative filtering and content-based filtering
industrial opportunity. Yet, the variety of clothing alternatives [1,2]. The first method uses the past interactions between users
on e-commerce platforms has posed additional difficulties for and items, such as how users have rated items in the past. The
consumers in picking the right outfit. Hence, in this project, we second method uses user profiles and item descriptions to
proposed a customized Fashion Recommender system that make recommendations. In recent years, collaborative
creates suggestions for the user depending on their input. This filtering systems have been using a Deep Learning-based
project aims to create product recommendations based on a Neural Collaborative filtering framework[4] that generalizes
user-provided picture of a product, since consumers often see the matrix factorization method. Modern recommendation
something, they like and then search for similar things. We use systems take into account things like a user's past purchases,
neural networks to analyze the Fashion Product Image Kaggle their feedback, the features of a product, when it was bought,
dataset's images, and a nearest neighbor algorithm to generate etc. But the way things look is an important factor that isn't
the final recommendations. taken into account by the current ranking and recommendation
systems. [15] added visual cues to the users' opinions and
Keywords— CNN, Fashion, recommendation, ResNet50 proposed a scalable factorization model that could be used on
large datasets from the real world. Most e-commerce sites get
I. INTRODUCTION (HEADING 1) their recommendations from the keyword mapping and the
With the emergence of recommender systems in a number knowledge database. This didn't work because the way each
of fields, retail businesses are investing in the latest buyer and seller described the item was different. General
technology to help their businesses grow. Fashion has been Recommender systems didn't do well when it came to
around for a long time, and it will still be around in the future. recommending fashion items because they were too
Women are more interested in fashion and style, and they have subjective. So, our approach uses image data of the item and
a wider range of items to choose from, which can make it hard shows that it is reasonable to rely on visual features to make
to decide. It has become an important part of modern families' item recommendations that are highly appealing and similar
lives because people are often judged by what they wear. Also, to the user's tastes and preferences.
clothing companies want their customers to try out all of their
products so they can choose what they like best. This can't be IV. PROPOSED METHOD
done by just going to a clothing store. The formatter will need We present a model that employs Convolutional Neural
to create these components, incorporating the applicable Network[5][6] and Nearest neighbor- algorithm supported
criteria that follow. With the fast expansion of Internet recommendation system. According to figure 1 Initially,
technology, intelligent clothing in the retail fashion business neural networks are trained, after which an inventory is chosen
has garnered more interest. The absence of suggestion for producing recommendations and a database for the
methods in the online shop annoys consumers who must do inventory's contents is constructed. The closest neighbor's
comprehensive searches. By adopting Convolutional Neural algorithm is utilized to identify the most relevant products
Network (CNN), we can able to distinguish similar clothing based on the input picture, and recommendations are then
and bring them into the front view based on previous data or provided.
picture input. This will help both the buyer and the seller,
allowing them to choose the perfect outfit and generate extra
sales. In this project, we built a novel, customized fashion
recommendation system based on the interests of the user. It
receives a single picture as input and produces a top-five
ranked list of clothing choices that are comparable to the
image.
II. DATASET
The dataset is acquired from Kaggle. The dataset includes
44,000 photos. Each product has a unique identifier, such as
42431. styles.csv includes a list of all of the items. File may

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

A. Trian the Model VI. DISCUSSION
Once the data has been preprocessed, neural networks are Due to the enormous compute power necessary for our
trained using ResNet50 transfer learning. In the last layers, situation, training the model takes too long. Aside from that,
further additional layers are added to replace the architecture owing to a lack of machine power, we were unable to test other
and weights of ResNet50 in order to fine-tune the network models for comparison. If sufficient machine power is
model for the present situation. Figure 2 illustrates the available, the end product will be even more precise.
architecture of ResNet50[7].
VII. CONCLUSION
In this project, we proposed a novel framework for fashion
recommendation that is data-driven, graphically connected,
and simple yet effective for producing photos of fashion
products. The proposed method consists of two phases.
Initially, our suggested method extracts the picture's
characteristics using a CNN classifier, e.g., enabling
consumers to submit any random fashion image from any
ecommerce website, followed by the generation of
B. Recommender Generation
comparable images based on the features and texture of the
To produce suggestions, we used the Python library input image. It is important that this kind of work continue, as
Streamlit. The closest neighbors are then shown in a column. it will help make fashion recommendations more precise and
We transformed the project into a web application to make it improve the whole fashion exploring experience for both
robust. The database's top 5 recommendations are retrieved, direct and indirect customers
and their photos are displayed [8].
REFERENCES
V. RESULT
The idea of transfer learning is utilized to overcome the [1] M. Sridevi and Dr. R.Rajeswara Rao A comprehensive Survey of
limitations of the fashion dataset. The dataset is used to train Personalized Recommender System Internation Journal of Computer
the classification model. The dataset is used to train and Science and Technology, Vol. 5 Issue 3, July-Sept 2014, ISSN: 0976-
8491.
validate the network. The training results demonstrate the
[2] M. Sridevi, Dr. R.Rajeswara Rao and Dr M.Varaprasad Rao, A Survey
model's high precision with minimal error, loss, and a high f1 on Recommender System 2016 Internaional Journal of Computer
score. Sciennce and Information Security, 14(5).
[3] X. He, L. Lio, H. Zhang, et al. Neural collaborative filtering. In
WWW2017, pages 173 182,2017.
[4] R. He and J. McAuley. Vbper: Visual byesian personalized ranking
from implicit feedback. In AAAI, Pages 144-150, 2016.
[5] Sallauddin, M. Rmesh, D. Harshvardhan, A. Pasha, S.N. Shabanan, “
A comprehensive study on traditionl AI and ANN architecture 2019
International Journal of Advanced Science and Technology 28 (17),
pp.479.
[6] Mahender K, Kumar TA and Ramesh KA PAPR 2018 analysis of fifth
genreration multiple access waveform for advanced wireless
These images, which are different in nature, are utilized to communication International Journal of Engineering and
test the proposed system. From the simulation of the Technology(UAE) 7(3.34 Special Issue 34) 487-490 10.214/NBHIV.
experiment's outcomes, we can conclude that despite relying S68956_old.
only on visual inputs, this strategy is robust and efficient. The [7] Sheshikala, M., Rao, D.R, & Prakash, R.V. (2017). Computation
graphic depicts the recommendation generated by our method Analysis for Finding Co-Location Pattenrs using Map-Reduce
for the picture input given. Framework. Indian Journal of Science and Techonlology, 10(8).
[8] Shesshikala, M,, 2019, Natural Language Processing and Machine
Learning Classifier used for Engineering (IJRTE).