Synthetic image generation for machine vision system training
About Project
Controlled generation of large volumes of images for training machine vision systems was required.
Value of the idea
Avoidance of privacy and security concerns
Data diversity
The generated images allow the creation of diverse datasets, which is especially important when there is a limited amount of real data.
Control of conditions
Quick data annotation
For some tasks, real data may contain sensitive information. Generated data solves this problem.
Scenario parameters including lighting, background, color, textures and other characteristics can be precisely specified, allowing you to test the behavior of the neural network under different conditions.
Since the generated data is already annotated, it reduces the time to prepare the data for training.
Creating images using current technology to train neural networks has several advantages:
Generated images of the construction process to train a neural network to detect safety violations
Complexities of the solution
In the process, problems arise that need to be solved for further use of the product:
- Unrealistic images (especially in medical graphics)
- Possibility of poor transfer to real data
- Incorrect annotations
For example, for realistic generation of synthetic documents (passport, driver's license, etc.), it is necessary to separate data and photos into male and female gender, as well as to take age into account during generation.
Generated passport with incorrect name. In some cases it is not necessary to take this parameter into account.
About Development
3D modeling was used to solve the problem.
Stages of Work:
1. Setting the overall composition of the scene. Creating the environment.
2. Modeling objects with different levels of detail.
3. Generating custom textures, applying them and procedural textures to the scene and objects.
4. Camera animation.
5. Image rendering and labeling using a specially designed script.
2. Modeling objects with different levels of detail.
3. Generating custom textures, applying them and procedural textures to the scene and objects.
4. Camera animation.
5. Image rendering and labeling using a specially designed script.
Areas of use
Synthetic image generation for training EO systems is a powerful tool and can have successful applications where required:
- Photo generation where it is difficult to get enough real images: confidential areas or the need to create rare or specific scenarios
- Generation under time and resource constraints
- Expanding the data set relative to existing data
Project team
Alexey V.
Developer
