Human-Guided Image Generation for Expanding Small-Scale Training Image Datasets.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Limited datasets hinder computer vision model performance in real-world applications like wildlife observation.
• Generative models can expand datasets but often lack diversity and control over output.
• Existing methods struggle with automatic, uncontrollable image generation, leading to undesired samples.

Purpose of the Study:

• To propose a human-guided method for controllable dataset expansion in computer vision.
• To enhance the diversity and relevance of generated images for improved model training.
• To address the limitations of automatic generative models in creating practical datasets.

Main Methods:

• Developed a multi-modal projection method for exploring original and generated images.
• Implemented a sample-level prompt refinement technique for user-guided feedback.
• Facilitated iterative refinement of image generation through user interaction and prompt adjustments.

Main Results:

• Quantitative evaluation confirmed the effectiveness of the multi-modal projection method.
• Improved performance of computer vision models in classification and object detection tasks.
• Expert feedback validated the usability and positive impact of the human-guided generation process.

Conclusions:

• Human-guided generation offers a controllable solution for dataset expansion in computer vision.
• The proposed methods enhance data diversity and model performance effectively.
• Sample-level feedback simplifies prompt refinement for improved image generation quality.