MAI CV
A set of hardware and software for detecting pipe defects.
At the implementation stage of the project, we achieved significant results over a one-month period, including conducting research and further development. The synthetic data generation tool we created can be used for related tasks of generating datasets and training
neural networks.
At the implementation stage of the project, we achieved significant results over a one-month period, including conducting research and further development. The synthetic data generation tool we created can be used for related tasks of generating datasets and training
neural networks.
Year of the project implementation: 2020
The client: MAI UAV Center
Duration: 1 month
The project team:
PM: Oleg Yusupov
R&D: Vadim Kondaratsev, Alexander Kruchkov, Roman Chumak
CG: Roman Chumak
DEV: Andrey Ivanov
The client: MAI UAV Center
Duration: 1 month
The project team:
PM: Oleg Yusupov
R&D: Vadim Kondaratsev, Alexander Kruchkov, Roman Chumak
CG: Roman Chumak
DEV: Andrey Ivanov
During the project implementation phase, we were able to achieve significant results over a month-long period, including research and further development. The synthetic data generation tool we created can be used for the interrelated tasks of generating datasets
and training neural networks.
A set of hardware and software for detecting pipe defects.
Reduction of time and financial costs for the maintenance of a thermal power plant. This process includes activities such as diagnosing pipes for damage or defects using unmanned aerial vehicles (UAVs). The data obtained is then analyzed using software. Our goal was to implement
a software prototype for the ability to detect defects based on photo and video data obtained by the UAV.
a software prototype for the ability to detect defects based on photo and video data obtained by the UAV.
We divided our project into several stages:
― Development of software for data labeling in CVAT;
― creating a tool for data augmentation;
― creation of a tool for generating synthetic data;
― training of neural networks (YoloV4, DetectoRS) for image analysis;
― creation of a client application that would serve to download a video stream from the UAV and analyze the condition of pipes.
― Development of software for data labeling in CVAT;
― creating a tool for data augmentation;
― creation of a tool for generating synthetic data;
― training of neural networks (YoloV4, DetectoRS) for image analysis;
― creation of a client application that would serve to download a video stream from the UAV and analyze the condition of pipes.
What we did and how we did it:
1. Research
Conducting an analytical report together with our partners from MAI (Moscow Aviation Institute) was a significant part of our project. We needed to prove that the goal could be achieved with the help of computer vision techniques, and choose the most effective way to implement the project.
1. Research
Conducting an analytical report together with our partners from MAI (Moscow Aviation Institute) was a significant part of our project. We needed to prove that the goal could be achieved with the help of computer vision techniques, and choose the most effective way to implement the project.
During the reporting phase, we did the following:
― Collect information about advanced image detection solutions;
― prepared and described 3 PhD deep architectures;
― described the current state
of datasets, data labeling tools,
and quality metrics in similar tasks
and projects;
― described options for using synthetic data generation tools.
― Collect information about advanced image detection solutions;
― prepared and described 3 PhD deep architectures;
― described the current state
of datasets, data labeling tools,
and quality metrics in similar tasks
and projects;
― described options for using synthetic data generation tools.
2. Development
At the next stage of our project, we implemented our own system for training deep architectures, data labeling, collecting logs and metrics, as well as visualization based on existing training systems. We created our own tool for generating synthetic data, which significantly accelerated the process of collecting information for training neural networks.
We then trained the selected architectures on different datasets (those that were collected together with our partners and those that we generated ourselves). As a result, we developed a prototype of a web client-server application, in which we integrated neural networks trained to detect pipe defects. The customer can upload data from the drone (in json format) to know both the visual and technical condition of the pipes.
2. Development
At the next stage of our project, we implemented our own system for training deep architectures, data labeling, collecting logs and metrics, as well as visualization based on existing training systems. We created our own tool for generating synthetic data, which significantly accelerated the process of collecting information for training neural networks.
We then trained the selected architectures on different datasets (those that were collected together with our partners and those that we generated ourselves). As a result, we developed a prototype of a web client-server application, in which we integrated neural networks trained to detect pipe defects. The customer can upload data from the drone (in json format) to know both the visual and technical condition of the pipes.
Learn more about the project on Behance:
Learn more about the project on Behance: