CNN Based Object Classification Using Image and Poisson Vector Field Features
Master of Science (MS)
Date of Award
Convolutional Neural Networks (CNNs) have great applications in image and video object recognition, classification and segmentation. In the literature on CNN models, their efficiency and efficacy are determined by their classification metrics, speed and computational cost. This thesis aims to improve the classification metrics accuracy of a CNN model by fusing vector fields’ features with image features. The vector field features of importance are the singular points and separatrices. These features are embedded in the training and testing images to increase the latter set’s classification metrics. This study applies particular vector fields developed on the solution of a specific form of the Poisson equations. These vector fields have both real and complex singularities. Then the vector fields are embedded into an image database to create a fusion of the vector fields’ features and the image databases’ features. The main novelty of the thesis is the development of a new CNN optimized to efficiently classify image databases with embedded vector field features. The goal is to validate that the embedding in question brings the advantage of increasing the classification statistics. The conducted experiments with the public skin lesions database ISIC 2020 successfully confirmed the claim.
Nikolay M. Sirakov
Applied Mathematics | Physical Sciences and Mathematics
Igbasanmi, Oluwaseyi, "CNN Based Object Classification Using Image and Poisson Vector Field Features" (2023). Electronic Theses & Dissertations. 1075.