Convolutional neural networks as feature extractors for data-scarce visual searches
Loading...
Authors
ben Abdallah, Hichem
Subjects
Convolutional Neural Networks
k-Nearest Neighbors
image classification
data scarcity
transfer learning
activation codes
high-dimensional space
cosine similarity
Euclidean distance
t-SNE
k-Nearest Neighbors
image classification
data scarcity
transfer learning
activation codes
high-dimensional space
cosine similarity
Euclidean distance
t-SNE
Advisors
Kolsch, Mathias
Date of Issue
2016-09
Date
Sep-16
Publisher
Monterey, California: Naval Postgraduate School
Language
Abstract
Image classification is one of the core problems in Computer Vision. The classification task consists of predicting a single label from a fixed set of categories for a single image. To perform an image classification, the classifier should consider the semantic identity of the image rather than irrelevant characteristics and variations such as the coincidental contrast or brightness of the images or the type of background. Applying Convolutional Neural Networks (CNNs) as feature extractors is a powerful approach to image classification. Training these CNNs necessitates a tremendous amount of training samples, and it is costly in terms of computational time. Since it is not guaranteed that one can find a sufficient amount of training data for a specific class target, we are conducting transfer learning of a CNN model (learned from a large data set) to generate a new representation of the images. These representations are classified with K-Nearest Neighbors within a target space that has just a few training samples.We aim to define the appropriate parameters including distance metric, layer from which to extract features, and minimum number of training samples to be considered to obtain the best classification results with our approach.
Type
Thesis
Description
Series/Report No
Department
Information Sciences (IS)
Organization
Identifiers
NPS Report Number
Sponsors
Funder
Format
Citation
Distribution Statement
Approved for public release; distribution is unlimited.
Rights
Copyright is reserved by the copyright owner.