Computer Vision Dataset Synthesis Using LoRA Tuned Latent Diffusion Models
Presentation Type
Poster
Student
No
Track
Other
Abstract
This research explores the potential of generative artificial intelligence to augment datasets for improving computer vision applications in specialized fields. Current challenges in dataset creation, such as manual image production, limit the availability of training data, impacting the performance of machine learning models that rely on large datasets for optimal accuracy.
To address this, a generative diffusion model fine-tuned using Low-Rank Adaptation (LoRA) is being developed to synthesize additional training data. The approach leverages an existing dataset as a foundation and generates synthetic images tailored to the target domain. These images are intended to complement the original dataset for training computer vision systems to improve classification accuracy in domain-specific tasks.
Initial experiments will involve training and evaluating multiple computer vision models using both the original and augmented datasets. Key performance metrics such as accuracy, loss, and F1 scores will be analyzed, alongside the impact of different model configurations, including activation functions and optimizers.
Preliminary findings aim to evaluate whether synthetic data can mitigate the limitations of small datasets, improve model performance, and demonstrate the feasibility of this approach for addressing data scarcity in niche research areas. The project’s outcomes are expected to contribute to the broader application of generative AI in improving machine learning workflows for specialized domains.
Start Date
2-7-2025 1:00 PM
End Date
2-7-2025 2:00 PM
Computer Vision Dataset Synthesis Using LoRA Tuned Latent Diffusion Models
Volstorff A
This research explores the potential of generative artificial intelligence to augment datasets for improving computer vision applications in specialized fields. Current challenges in dataset creation, such as manual image production, limit the availability of training data, impacting the performance of machine learning models that rely on large datasets for optimal accuracy.
To address this, a generative diffusion model fine-tuned using Low-Rank Adaptation (LoRA) is being developed to synthesize additional training data. The approach leverages an existing dataset as a foundation and generates synthetic images tailored to the target domain. These images are intended to complement the original dataset for training computer vision systems to improve classification accuracy in domain-specific tasks.
Initial experiments will involve training and evaluating multiple computer vision models using both the original and augmented datasets. Key performance metrics such as accuracy, loss, and F1 scores will be analyzed, alongside the impact of different model configurations, including activation functions and optimizers.
Preliminary findings aim to evaluate whether synthetic data can mitigate the limitations of small datasets, improve model performance, and demonstrate the feasibility of this approach for addressing data scarcity in niche research areas. The project’s outcomes are expected to contribute to the broader application of generative AI in improving machine learning workflows for specialized domains.