DETECTION OF DENSE, OVERLAPPING, GEOMETRIC OBJECTS.

Area of Science:

• Computer Vision
• Data Visualization
• Materials Science

Background:

• Scientific publications often use scatterplots to display thermophysical property data.
• These plots contain geometric objects (circles, triangles, squares) that require accurate detection and classification.
• Varying object density, clarity, and size present challenges for automated image analysis.

Purpose of the Study:

• To develop and evaluate a deep learning model for detecting and classifying geometric objects in scatterplot images.
• To investigate the impact of training data annotation strategies on model performance for object localization and classification.
• To optimize image analysis techniques for scientific data visualization.

Main Methods:

• A unique dataset of black and white scatterplot images from scientific journals was compiled.
• A single-class U-Net convolutional neural network was designed and trained for object detection.
• Different annotation patterns for training data masks were explored to assess their effect on accuracy.

Main Results:

• The U-Net model achieved 97% accuracy in identifying image objects and locating their centers within a few pixels.
• Specific annotation strategies for training data masks significantly improved both object classification and localization accuracy.
• Optimizing annotations proved more impactful than modifying network loss terms for performance enhancement.

Conclusions:

• Deep learning, specifically U-Net models, can effectively detect and classify geometric objects in complex scientific image data.
• The method of annotating training data masks is crucial for achieving high accuracy in both object localization and classification.
• Distinct annotation requirements exist for optimizing point localization versus geometric object classification.