Selection of universal features for image classification

Rodríguez, Pedro; Drenkow, Nathan; DeMenthon, Daniel; Koterba, Zachary H.; Kauffman, Kathleen; Cornish, Duane C.; Paulhamus, Bart L.; Vogelstein, R. Jacob

doi:10.1109/wacv.2014.6836078

Cited by 2 publications

(2 citation statements)

References 12 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Conceptually, an important feature is one that is able to cause large change in classification performance; for example, [17] proposes a scheme to measure the importance of features. Thus, further work will focus on the selection of important features, which will reduce the amount of relevant data.…”

Section: Conclusion and Further Workmentioning

confidence: 99%

“…The overall performance of object-specific features is better than that of the universal features, while the latter becomes competitive for smaller training sets. In another natural image processing paper [17], a method to calculate the importance of universal patches is proposed, and good image classification performance can be achieved by using a relatively low number of highly important patches. To the best of our knowledge, this aspect of universality to a set of supervised learning problems has not been previously formalized.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Universality of wavelet-based non-homogeneous hidden Markov chain model features for hyperspectral signatures

Feng

Duarte

Parente

2015

2015 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

View full text Add to dashboard Cite

Feature design is a crucial step in many hyperspectral signal processing applications like hyperspectral signature classification and unmixing, etc. In this paper, we describe a technique for automatically designing universal features of hyperspectral signatures. Universality is considered both in terms of the application to a multitude of classification problems and in terms of the use of specific vs. generic training datasets. The core component of our feature design is to use a non-homogeneous hidden Markov chain (NHMC) to characterize wavelet coefficients which capture the spectrum semantics (i.e., structural information) at multiple levels. Results of our simulation experiments show that the designed features meet our expectation in terms of universality.

show abstract

Section: Conclusion and Further Workmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%