Whole Image and Modular Image Face Classification - What is Really Classified?

摘要

Our aim is to explore the importance of chosen parts of frontal face images for person recognition. We have used logistic regression as the method of face image classification based on rough image classification, and on selected parts of an image divided into rectangular image blocks. Rough image means that no image processing transformation is performed before classification. Experiments on the images of 40 persons taken from the ORL face database show that a person classification based on collections of rough face images are effective, high accuracy rates are easy to obtain, but deeper analysis based on image partitioning suggests that the most important factor for correct classification are border parts of the face image. Furthermore, the experiments confirm the thesis that randomly generated projections do not degrade, or only slightly reduce the accuracy of classification, reducing the size of the vector of features in a significant way.