Encoding the SIFT Features with BoW Framework

In this study, we exploit the BoW framework to represent the staining patterns of HEp-2 cells. The BoW frameworkhas been introduced with detail in Chap. 4. Briefly, we firstly divide the HEp-2 cell image into small overlapped patches. Secondly, we extract the SIFT features which are invariant to scaling and rotation, and partially invariant to illumination change, viewpoint change and noise. These properties are advantageous in staining pattern classification as most of HEp-2 cell images present large orientation variations and have high intra-class variability. The patch-level SIFT features are extracted from patches within an image, i.e., X = [x1; x2xN] e RDxN . N is the number of patches of the image and the dimension of SIFT features is 128.

Thirdly, all the patch-level SIFT features of the images for training are clustered to generate a codebook B = [b1, b2, ••• , bM] e RDxM. Generally, a larger number of visual words can achieve a higher classification accuracy but needs higher computational cost.

Then, by using a specific coding method, SIFT features of each image can be transformed to a collection of feature codes, i.e., C = {c1; c2,..., cN} e RMxN. Finally, the image is divided into increasingly finer spatial regions. Multiple codes from each subregion are pooled together by the max-pooling strategy to retain spatial information of the visual words’ location. The final representation is generated by concatenating the histograms from all subregions together.

< Prev   CONTENTS   Source   Next >