sentin GmbH

Brüderstraße 5
44787 Bochum

Is your company ready for AI &

Machine Vision projects?
The 24 questions you should ask yourself.

When it comes to innovation projects there are several things to consider. To save
you time and effort we curated this questions and examples. We divided the
questions in 4 categories focusing on the steps we believe are most crucial.

If the wording feels unfamiliar to you, you can read more about

- Artificial Intelligence, Deep Learning

- Training & Models

- Classes/Features

in this FAQ

sentin GmbH | Brüderstraße 5, 44787 Bochum | HRB 17706 | USt-IDNr. DE322903907

Deutsche Bank Bochum, IBAN DE30430700610234500700, BIC DEUTDEDE430
The Checklist – 4 Categories
We are using 4 major categories to determine if you are ready for AI projects. They
contain important questions about the requirements of your current processes, the
technical, financial and project related topics. We summed them up as the
following:

1. The Status Quo - Current task, inspection or evaluation process

2. The Data Requirements
3. Your return on investment (ROI) requirements
4. Bringing the system to production

Part 1 - The Status Quo

The current task, inspection or evaluation process.

You are the expert of your inspections or evaluations. This section helps to translate
your requirements to the AI world and can help to identify the right AI architecture.
The following questions help to identify how complex the evaluation is and what
classes/features/defects should be considered.

[ ] 1 of 4 - Do you have objective criteria for the evaluation?

This could be a norm or quality guideline ( e.g. if a crack is immediately a defect, but
porosities <5mm are ok ). This helps to qualify the results and the requirements of
the Artificial Intelligence.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 2 of 4 - Do you care for a defect or feature dimensions/count or

location?
If you are interested in the size of a crack or you just want to count something in an
image this is relevant for the AI model architecture and output.

___________________________________________________________________________
___________________________________________________________________________
___________________________________________________________________________
___________________________________________________________________________

sentin GmbH | Brüderstraße 5, 44787 Bochum | HRB 17706 | USt-IDNr. DE322903907

Deutsche Bank Bochum, IBAN DE30430700610234500700, BIC DEUTDEDE430
[ ] 3 of 4 - How many different features do you want to classify/detect?
If you check for cracks, malformed, porosity and incompletion this would be 4. If
you just decide between OK and Not Ok this would 2 (called binary classification).

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 4 of 4 - How many different parts, views or locations do you check?

AI models can be very robust, but if you have many different parts and views e.g. on
3 production lines you should consider this in the requirements.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

Part 2 - The Data Requirements

The more data the better.

As a rule of thumb: the more complex or detailed the task the more data you need.

The more you know of your data the better an AI can be trained. If your images are
digital and well-organized AI experts can work on solutions faster and achieve
better results.

Did you know 80% of an AI project is about data preparation and organization?

[ ] 1 of 5 – Do you know how many digital images do you have?

The training of an AI solution needs data. There are rare cases where you can use
an out of the box model. If your inspection is common, you can reduce the number
of images needed, but if it is unusual or has a lot of variety the amount may increase.
What is the distribution between the different classes. Do you have 100 images of
cracks, but only 10 of porosities? This might be a problem?

If you do not have any digital images, is there a way to digitize or collect them?

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

sentin GmbH | Brüderstraße 5, 44787 Bochum | HRB 17706 | USt-IDNr. DE322903907

Deutsche Bank Bochum, IBAN DE30430700610234500700, BIC DEUTDEDE430
[ ] 2 of 5 - Which file type are you using? Are they compressed?
We recommend using TIFF or PNG files, because they can be saved lossless. This
can be better, because the AI model can analyze more information. JPGs are
smaller but may loose some image information due to their compression?

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 3 of 5 – Do you know your average file size? What is the minimum /

maximum?
The larger the image, the more information has to be processed. This can have an
impact on the evaluation time? Some companies check really large parts and
generate huge files (e.g. 10.000 x 10.000 pixel - 100 Megabyte). Others may check
small parts with small images (1024 x 1024 pixel - 500 Kilobyte). This also has an
impact on the hardware you need to train or evaluate.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 4 of 5 – Do you have your data stored online?

If your images are on an analog film in your basement, it could take some time to
use them for AI training. Better would be a digital version on a SSD. We recommend
a cloud connection because of the backup capability and availability.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 5 of 5 – Do you have labelled/annotated data?

This could be a simple sorting per file system or filename. One folder with defects
and one folder with passed images. Alternatively there are tools that allow you to
manage your files for AI projects and label features inside the images.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

sentin GmbH | Brüderstraße 5, 44787 Bochum | HRB 17706 | USt-IDNr. DE322903907

Deutsche Bank Bochum, IBAN DE30430700610234500700, BIC DEUTDEDE430
[ ] If you have labelled data, did you use a special tool? Which format have you used
e.g. COCO format. (optional)

__________________________________________________________________________________
__________________________________________________________________________________

[ ] If not, can you label them or know someone who can? (optional)

__________________________________________________________________________________
__________________________________________________________________________________

Part 3 - The Return On Investment (ROI)

How is the current process affecting your business?

We have talked about the technical requirements for an AI project. The following
questions deal with the financial part and can help to convince decision makers.
The questions can also help to make a first calculation for the ROI - coming from
process cost and time saving.

[ ] 1 of 8 – Do you know how many images you process per week?

Normally this depends on the inspection or evaluation. We know projects where
there were about 5.000 images per week e.g. in non-destructive testing of
pipelines. Other projects had 20 Mio. images per week e.g. in the public sector and
others had about 100 per week for large parts.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

sentin GmbH | Brüderstraße 5, 44787 Bochum | HRB 17706 | USt-IDNr. DE322903907

Deutsche Bank Bochum, IBAN DE30430700610234500700, BIC DEUTDEDE430
[ ] 2 of 8 – Do you know the average image evaluation time? Are there
any limitations?
In non-destructive testing there are cases that take up to 7 minutes per image or
longer. In manufacturing you need fast processing. If you take the number of
images per week by the time it takes you to inspect a single one, you can get an
indicator of time saving. For most cases AI models can evaluate images in less than
a second or even in a few milliseconds.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 3 of 8 – How many people/machines are doing the inspection?

In some branches it is hard to find qualified personnel. If you could save the time of
your inspectors, they could focus on other tasks. If you already have an inspection
system or automation solutions in place, are they performing well?

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 4 of 8 – Do you know your current accuracy? What is the desired

accuracy?
Quality controls or inspections need high detection rates. If you want to use AI, you
need to know where you are at the moment and what you are aiming at. Otherwise
it will be hard to determine if the AI solution really performs better.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 5 of 8 – Do you want the evaluation to be more reliable or do you want

to reduce false alarms?
This question determines what the focus of your evaluation should be. Some
people come from a low detection rate resulting in many false-negatives others
come from high false-positives. In manufacturing this often means pseudo rejects
(parts that are actually good, but falsely accused of being defect) or high customer
complaints (because the system failed to detect parts that were defect).

__________________________________________________________________________________
__________________________________________________________________________________

sentin GmbH | Brüderstraße 5, 44787 Bochum | HRB 17706 | USt-IDNr. DE322903907

Deutsche Bank Bochum, IBAN DE30430700610234500700, BIC DEUTDEDE430
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 6 of 8 – Do you know what a wrong evaluation costs?

For public sector this can be a fine or if an inspection is wrong people can be hurt.
In manufacturing this could be pseudo scrap or customer complaints. There are
projects where a single complaint (of a plastic part) costs 1.000€ with all the internal
costs.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 7 of 8 – Do you have a clear goal for using an AI system?

This question might sound trivial, but it can help to identify if you even need an AI
system. Rule-based Machine Vision systems (not AI) have been around for some
time. If you just want to check if something has passed an area a photoelectric
sensor might be enough. Not every task needs an AI to perform well.

The main benefits AI system offer are that they get better and more robust the
more data they have and that they can work with complex tasks. Then they can be
faster, cheaper and more accurate as other solutions.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 8 of 8 – Do you have run prior AI or Computer Vision projects?

Artificial Intelligence is coming to more and more companies. Some companies
have been early adopters. But most need to have tested and reliable solutions. If

sentin GmbH | Brüderstraße 5, 44787 Bochum | HRB 17706 | USt-IDNr. DE322903907

Deutsche Bank Bochum, IBAN DE30430700610234500700, BIC DEUTDEDE430
you already have run automation projects or have an computer vision engineer in
your team this could be a benefit.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

Part 4 – Bringing a system to production

The key to successfully using an AI.

The last part is about the hurdles that can make it hard to complete a project
successfully. It is dealing with the organization and deployment of the solution.
Many projects fail because this questions do not get asked or get asked too late.

[ ] 1 of 7 – Do you have the right hardware (e.g. GPUs)? Who is responsible

for maintaining it?
To successfully run an AI model you may need special hardware. The computation
power needed is different from normal desktop applications, because they process
image data. If something has to be bought or installed, who is responsible for that?

__________________________________________________________________________________
__________________________________________________________________________________

sentin GmbH | Brüderstraße 5, 44787 Bochum | HRB 17706 | USt-IDNr. DE322903907

Deutsche Bank Bochum, IBAN DE30430700610234500700, BIC DEUTDEDE430
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 2 of 7 – Do you have internet or network access?

Cloud applications are the future of software, but not every location has internet
access. If you have mobile inspection team you may depend on LTE. Other
networks are encapsulated from the internet. This is something you have to
consider, because where is the computation power and how do updates come to
the system.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 3 of 7 – Do you have a project responsible? Who supports it?

Innovative projects need the support of the participants. If the CEO understands
that this project will boost the company's efficiency with a clear ROI it may be easier
to get a budget. If the IT department understands the requirements of the project,
it will be faster to install the needed hard- and software. Therefore, be sure to have
all information and requirements.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 4 of 7 – Do you have the financials clear?

Before you start to set up an AI project you should ask yourself which department
will pay for it? Is there a budget for innovation projects? Is there a budget for solving
the process problem? Who needs to be involved in the process to make the buying
decision? Often it saves a lot of time if the buying center knows right from the start
that there may be a new supplier.

sentin GmbH | Brüderstraße 5, 44787 Bochum | HRB 17706 | USt-IDNr. DE322903907

Deutsche Bank Bochum, IBAN DE30430700610234500700, BIC DEUTDEDE430
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 5 of 7 – Do you have a timeline for the project?

The questions before should have shown, that there are many things to consider. If
you do not have any images or all of them are analog this will take longer than
boosting the accuracy of an existing machine vision system.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 6 of 7 – Do you know how to verify the results of an AI?

This is one of the most important questions for a successful AI project. When is the
AI better than your prior solution? Better accuracy? Faster evaluation? If this Is not
answered you may struggle to really use the AI model and stick to the old solution.

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

[ ] 7 of 7 – Do you know how to update an AI system?

What happens if your process changes? How do you adapt? Can you update your
system on your own? Who does the maintenance?

__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________
__________________________________________________________________________________

sentin GmbH | Brüderstraße 5, 44787 Bochum | HRB 17706 | USt-IDNr. DE322903907

Deutsche Bank Bochum, IBAN DE30430700610234500700, BIC DEUTDEDE430