Locally Linear Embedding based on Image Euclidean Distance

Abstract

We present an improved Locally Linear Embedding algorithm based on Image Euclidean distance (IMED) to replace the traditional Euclidean distance. IMED depending on pixel distance is robust to the noises in images. So in theory, applying the new distance metrics to LLE can bridge a gap, that is, traditional LLE is sensitive to noises. The improved algorithm highly enhances its stability to noises. We apply the algorithm to face detection, with SVM as the classifier, in the CBCL face database and test the detector on CMU frontal face test set. The result demonstrates a consistent performance improvement of the algorithms over the original version.