A system for detecting foreign objects on the underside of cars using a neural network
Image analysis
Basic information
Customer
ISS - Intelligent Security Systems - company-integrator of security solutions
Product
Task
Software for test benches of a car - the underside is inspected by a person using mirrors and photographs
Find abnormal objects at the bottom of the car using a photo
Where this can be used:
at military sites
Why it is needed:
at border posts
in places with increased safety requirements
automate the verification process
speed up the operator's work
How it works
The car arrives at the stand
The car's underside is photographed
When anomalies are found, it shows the snapshot to the operator
The neural network analyzes the snapshot
Anomaly
Technical implementation
Conducted ML audit
Since such tasks were almost not solvable, our client ordered an ML-audit service
Found a scientific article
As a result of the research, we found a scientific article on a similar topic.
Model
We reproduced the work detailed in the article in order to adapt our own data to the model
VisionHub
Using our VisionHub machine models we implemented our visualization interface
GAN
The main method is generative adversarial networks (GAN), a machine learning algorithm.
Principle of operation:
Step 01
Remember how an object looks like in its normal state
Step 02
Generate images and capture anomalies on them
Anomaly detected
Development process
Metric - the ratio of correctly identified anomalies to false ones
Anomaly: yes / no on this fragment. There was no task to accurately localize.
Anomaly: yes / no on this fragment. There was no task to accurately localize.
Input data
20
6
images of the underside of cars with anomalies
reference images of the bottom
Very small dataset, insufficient for high-quality neural network training
High resolution images - 4K by 10K pixels
Dataset processing
24
317
We manually marked the images, noted the anomalies.
images with anomaly
normal
We cut the photos into 512-pixel squares and looked at them separately.
130
3221
abnormal
normal
Model training
Trained the model on a prepared sample.
The accuracy of the quality metric was 0.73
The accuracy of the quality metric was 0.73
Development stages
Harmonization
Audit
Start of development
1.5-2 months
2 weeks
Current results
The audit stage has ended
Managed to build the first model
Getting ready for a big project
Project team
Manager
Mikhail Evdokimov
Tech lead
Vlad Vinogradov
Computer Vision Engineer
Aelita Shaikhutdinova
Features of the project
First used GAN in production
One of our first audits
This is the first time we work with an integrator company and, in general, with people who are very indirectly related to the development.
Where groundwork can come in handy
For example, in medical applications
In systems where deviation from the norm needs to be detected
Difficulties of the project
Unusual material
Communication difficulties, since the company is very large
Contact us
PRODUCTS
SECTIONS
STUDIOS
Russia, Innopolis, st. Universitetskaya, 7
2023 © EORA. All right reserved