6 April 2016 Fourier spatial frequency analysis for image classification: training the training set
Author Affiliations +
Abstract
The Directional Fourier Spatial Frequencies (DFSF) of a 2D image can identify similarity in spatial patterns within groups of related images. A Support Vector Machine (SVM) can then be used to classify images if the inter-image variance of the FSF in the training set is bounded. However, if variation in FSF increases with training set size, accuracy may decrease as the size of the training set increases. This calls for a method to identify a set of training images from among the originals that can form a vector basis for the entire class. Applying the Cauchy product method we extract the DFSF spectrum from radiographs of osteoporotic bone, and use it as a matched filter set to eliminate noise and image specific frequencies, and demonstrate that selection of a subset of superclassifiers from within a set of training images improves SVM accuracy. Central to this challenge is that the size of the search space can become computationally prohibitive for all but the smallest training sets. We are investigating methods to reduce the search space to identify an optimal subset of basis training images.
© (2016) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
Timothy H. Johnson, Yigah Lhamo, Lingyan Shi, Robert R. Alfano, Stewart Russell, "Fourier spatial frequency analysis for image classification: training the training set", Proc. SPIE 9711, Imaging, Manipulation, and Analysis of Biomolecules, Cells, and Tissues IX, 97110T (6 April 2016); doi: 10.1117/12.2212934; https://doi.org/10.1117/12.2212934
PROCEEDINGS
5 PAGES


SHARE
Back to Top