![](bg2.png)
and dispatch (Alan et al., 2014). However, these
types of solutions are very general and solve partially
the common problems, especially those presented by
stockists when relocating products that are in incor-
rect locations or verifying the status of a large number
of products that are available to the customer.
That is why many companies, including retailers,
have begun to invest in the use of Artificial Intelli-
gence (AI) to develop solutions to these problems.
According to new research from Juniper Research,
global spending by retailers on AI-enabled services
will reach $12 billion by 2023. Retailers’ use of AI
will make back-office operations more efficient. Fea-
tures such as demand forecasting and automated mar-
keting, under the influence of AI, will allow retail
businesses room for improvement and become more
agile. In addition, it is predicted that there will be a
dispute among retailers to include AI in their activi-
ties first. Those that include it will displace those that
do not have it implemented. As a consequence of its
implementation, the services offered will be superior
and prices for customers will be optimized
3
.
These types of investments are related to the de-
velopment of AI-based applications, which have been
increasing in recent years. It is increasingly common
to see customers in supermarkets scanning products
with their mobile phones to analyze them before buy-
ing them. These applications use computer vision
models to offer different functionalities. The factor
of automating processes thanks to computational vi-
sion is very important when developing these applica-
tions. There are, for example, applications for auto-
matic image retouching. Clothing companies receive
thousands of unique items that must be processed into
a final product that is professional and appeals to buy-
ers. This means that each image of each product must
be classified and labeled. Such a process is quite ex-
pensive and prone to error if done by a person. Au-
tomating the process of retouching one of these im-
ages using computer vision can take up to 30 times
less than if it were done by a professional.
Applications such as SolidGrids allow, through
computer vision, to automatically retouch, sharpen
and eliminate the background of the image of the de-
sired product. Other types of applications are those
that recommend products by visual similarity. This
can be very useful, not only to be able to navigate be-
tween the different items in the catalogue, but also
to solve the problems that the lack of stock of the
first chosen product would generate. Each product
3
“AI Spending by Retailers to reach $12 billion by 2023,
driven by the Promise of Improved Margins” - https://ww
w.juniperresearch.com/press/ai-spending-by-retailers-rea
ch-12-billion-2023
can be represented under its attributes and a category
to which it belongs, to perform, for example, filters
that the customer requires when looking for a type
of product, but without having a description or la-
bel (Santra and Mukherjee, 2019) .
Under the premise of this last class of application
in the retail sector and in order to solve part of the
problem that is stock management by storekeepers,
we developed a user-friendly mobile app for stock-
ists themselves. Through the training of the YOLOv5
object detection model, the model that will allow de-
tecting, counting and detecting the status of products
in an image was developed. Our work is limited to
the detection of canned and bottled products through
a photo taken or uploaded, within the context of cur-
rent Peruvian supermarkets. Furthermore, detecting
the current condition of the product simply indicates
whether the product is in good or bad condition. The
contributions of our work are the following:
• We implement the YOLOv5 object detection
model for localization and state detection of bot-
tled and canned products.
• We developed our own canned and bottled prod-
uct image dataset for training the object detection
model YOLOv5.
• We developed a mobile application for the man-
agement of products in supermarkets, which al-
lows daily records of those products that are de-
tected by uploading a photo.
In Section 2, similar works to ours are discussed. In
Section 3, the main notions required to develop our
work are detailed, and the main contributions of our
work. In Section 4, all the experiments carried out
are described to prove the feasibility of our proposal.
Finally, Section 5 present the main conclusions.
2 RELATED WORKS
Next, a brief description will be made on different
works and existing solutions for the product recog-
nition with different technologies. Additionally, solu-
tions were found that seek to detect the status of cer-
tain products, similar to the purpose of our proposal.
In (Selvam and Koilraj, 2022), the authors pro-
pose a framework for retail product detection consist-
ing of three modules: Product Detection, Product Text
Detection, and Product Recognition. For product de-
tection it uses the YOLOv5 model. To improve the
performance of the “TextSnake” algorithm in the sec-
ond module by replacing the backbone and using the
WHBBR (Width Height based Bounding Box Recon-
struction) processing technique in order to detect reg-
ICSBT 2023 - 20th International Conference on Smart Business Technologies
84