Interactive Visual Study of Multiple Attributes Learning Model of X-Ray Scattering Images

Area of Science:

• Materials Science
• Computer Science
• Scientific Visualization

Background:

• Interactive visualization tools for deep learning primarily focus on natural images, leaving a gap for scientific applications like x-ray image classification.
• Analyzing multiple structural attributes in x-ray scattering images presents unique challenges for current deep learning models.

Purpose of the Study:

• To present an interactive system for domain scientists to visually study multi-attribute learning models applied to x-ray scattering images.
• To enable exploration of scientific images within model-defined embedded spaces, facilitating attribute analysis.

Main Methods:

• Developed an interactive system allowing visual exploration of x-ray scattering images.
• Integrated model prediction outputs, actual labels, and neural network feature spaces for analysis.
• Enabled flexible selection and comparison of image instances and clusters based on attribute representations.

Main Results:

• The system facilitates visual study of attribute relationships impacting model accuracy.
• Users can identify and address issues like questionable attribute labels and outlier data clusters.
• Demonstrated improved training datasets and model refinement through interactive exploration.

Conclusions:

• The developed system enhances the understanding and improvement of deep learning models for x-ray scattering image classification.
• It empowers domain scientists to refine models and datasets by visually inspecting attribute learning.
• Case studies confirm the system's utility and effectiveness in scientific image analysis.