Automatic Extraction and Classification of Flow Visualization Features using Machine Vision
Faculty Mentor Name
J. Matt Pavlina, Akhan Almagambetov
Document Type
Presentation
Location
Jim and Linda Lee Planetarium
Start Date
4-10-2019 3:30 PM
End Date
4-9-2019 3:40 PM
Abstract
This research project is a preliminary investigation into the automatic analysis and classification of wind tunnel flow visualization images. Currently, images obtained during wind tunnel testing must be categorized manually by an individual. This type of classification requires years of experience and time. The first aspect of the project is to generate the most accurate vector representation of the lines in the flow visualization image. Machine vision is used to make these patterns interpretable by a computer, and neural networks are used for fine-tuning. Computer vision techniques are implemented to automatically skeletonize smoke flow visualization images with an algorithm that segments and scans along the individual smoke bands within the image. Two different methods are compared: a left-to-right scanning method, and a second method based on the binary representation of the image. Throughout the process, thresholding numbers are selected manually and vary based on the image being analyzed. To maintain autonomy, a classification neural network is introduced to automatically select these thresholding parameters for all images. With more accurate automatic vector representations of flow visualization images, images will be able to be analyzed further with machine learning techniques to extract useful information from the images.
Automatic Extraction and Classification of Flow Visualization Features using Machine Vision
Jim and Linda Lee Planetarium
This research project is a preliminary investigation into the automatic analysis and classification of wind tunnel flow visualization images. Currently, images obtained during wind tunnel testing must be categorized manually by an individual. This type of classification requires years of experience and time. The first aspect of the project is to generate the most accurate vector representation of the lines in the flow visualization image. Machine vision is used to make these patterns interpretable by a computer, and neural networks are used for fine-tuning. Computer vision techniques are implemented to automatically skeletonize smoke flow visualization images with an algorithm that segments and scans along the individual smoke bands within the image. Two different methods are compared: a left-to-right scanning method, and a second method based on the binary representation of the image. Throughout the process, thresholding numbers are selected manually and vary based on the image being analyzed. To maintain autonomy, a classification neural network is introduced to automatically select these thresholding parameters for all images. With more accurate automatic vector representations of flow visualization images, images will be able to be analyzed further with machine learning techniques to extract useful information from the images.