Development and Deployment Project
Uploaded by
ScribdTranslationsDevelopment and Deployment Project
Uploaded by
ScribdTranslationsSubject student_data Date
Research in Acosta Vela
22/05/2023
Artificial Intelligence Ruben Eduardo
Activity: Proposal for development and deployment of
an artificial intelligence project
System requirements: what need do we want to address?
In supermarket cash registers, a table with codes is used that
relate the price per gram with the names of fruits and vegetables. When one is
when purchasing these types of products, the cashier must register them one by one
codes, which delays the payment process. The idea of this project is to include a
camera with an automatic recognition system that records the product that is
about the scale and I identified it to avoid that work for the cashier and thus speed up the
product registration.
Solution design
We need to build a recognition model for fruits and vegetables that is very
versatile because on the scale you don't just put a lemon or a tomato for example, but
a bag full of lemons, with different sizes and even colors, and that makes it difficult to
correct identification.
To have a robust model we need to have a large set of images, with
various examples for each product in terms of size, shape, ripeness, and type. The
The main source of this data will be public datasets obtained from sites such as
Kaggle, data.mendeley, and Hugeface, but it should also be enriched with images.
taken under conditions similar to those of supermarket cash registers.
The model will be a classifier, but given the great diversity of elements, it is necessary
to study which method or set of methods generate better results. It is proposed
employ random forest methods, knn, and convolutional neural networks. Given the
The computational cost of neural networks will be compared in terms of
Research in Artificial Intelligence
1
Theme 7. Activities
Subject studentData Date
Research in Acosta Vela
May 22, 2023
Artificial Intelligence Ruben Eduardo
performance and resources used between simple machine learning algorithms and the
of Deep Learning.
In the real implementation of this system, it cannot be guaranteed that the conditions
of homogeneous lighting, nor the same quality of the images nor the
spatial distribution of the product within the image, which makes it essential a
preprocessing of the same before applying the classification algorithms for
optimize your results.
To correct lighting conditions, an equalization of the
histogram and the morphological closing filter that eliminates the specular reflection that
can generate some bright fruits. Then we need to segment the image to
obtain the edges of the fruits, create a binary mask with which to erase
the background of the image and center it. This final image is the one provided as input
to the classification model.
Development methodology.
A hybrid methodology between scrum and spiral is proposed. The objective of the part
Scrum is about planning intermediate objectives in each sprint that allow for control of the
progress of the project and have clarity on the stages of development that it requires,
both in data collection, its preprocessing, analysis, comparison of the
models and optimization of these, but a spiral part is required since each
result can feed back into the previous process and improve it, thus having a
iterative development, testing, and evaluation system.
How would you deploy the system?
The system must be lightweight to be run locally and in real time. It is discarded
a deployment in the cloud since it requires a stable internet connection and
Research in Artificial Intelligence
2
Theme 7. Activities
Subject student_data Date
Research in Acosta Vela
22/05/2023
Artificial Intelligence Ruben Eduardo
this cannot be guaranteed in many supermarket locations, moreover
that a loss of connection would disable the recognition system and with it the
product registration.
Each cash register has a computer, so the system of both
Preprocessing like the classifier model must be lightweight enough.
to be executed on these devices and thereby ensure access to the classifier.
Each cash register must also be equipped with a camera to take the
photograph that will be sent to the recognition system.
Research in Artificial Intelligence
3
Topic 7. Activities
