doi:10.1016/S0167-8655(99)00155-5 
Copyright © 2000 Elsevier Science B.V. All rights reserved.
Image classification using the frequencies of simple features
Clark S. Lindsey
, 
and Michael Strömberg
Department of Physics (Frescati), Royal Institute of Technology, Frescativägen 25, S-10405, Stockholm, Sweden
Available online 16 April 2003.
References and further reading may be available for this article. To view references and further reading you must
purchase this article.
Abstract
We investigate using the frequency of simple features to provide image signatures for input to a classifier. In an approach inspired by the n-gram technique for text classification, a binary image is scanned with a small window, e.g. 3 × 3 matrix and the occurrences of all possible features patterns within that window are counted. A vector with an element for each possible feature is then created with the element coefficients proportional to the frequency of the corresponding features or p-grams, e.g. the vector would have 512 elements for a 3 × 3 window. We tested the method by calculating the p-grams of artificially created images of four different objects and presenting them to a self-organizing map (SOM). We found this classification scheme successful for this limited image domain. The p-gram encoding scheme provides invariance to translation of the objects within the image and tolerance to scale variations as well.
Author Keywords: Image classification; Image encoding; SOM; p-Gram
Fig. 1. Illustration of p-gram construction with (a) a 3 × 3 filter and (b) an image matrix of binary pixels. The 3 × 3 filter in (a) is convoluted with the binary image matrix to count the number of times each of the 512 possible 3 × 3 binary patterns occur. For example, operation on the gray area in (b) would result in a VALUE =1 × 1 + 32 × 1=33, where only the non-zero multiplication is indicated. The 33rd element of the p-gram vector would thus be increment by one. The filter then slides by one column to the right and the operation repeats until reaching the right side of the image. Then the filter slides one row down and so forth until the whole image is scanned.
Fig. 2. Here are shown four types of artificial assembly images created to test the p-gram method with SOM categorization. The fuel rods are not perfectly round but instead have edges with random jaggedness. Also, random noise pixels were added and the assemblies were given random translations within the image.
Fig. 3. Outputs of an 8 × 8 self-organizing map for a test set of p-grams for 120 images of assemblies as shown in Fig. 2. There were 30 images for each of the four assembly types in the test set. Trained on a set of 1600 patterns, the test results here show that the SOM learned to separate cleanly the four assembly types: (+) BWR-L, (*) BWR-R, (×) BWR-X, (○) PWR . The symbols have been scattered around each point for the sake of showing the number of patterns.
Abstract
Purchase PDF (121 K)
E-mail Article
Add to my Quick Links
Cited By in Scopus (2)
Purchase PDF (1191 K)
