Robustness Evaluation of Extracting Features Based on Self-organizing Map Neural Network

Abstract

The descriptors should be robust to changes of images taken under various conditions in order to obtain correct recognition. HOG, LBP and convolutional neural network (CNN) are proved to perform better in extracting features of images. However, little attention was paid to the robustness of these methods. In this paper, we proposed a self-organizing map (SOM) neural network based descriptor evaluation method in order to assess the robust image features. It efficiently observes the features extracted by HOG, LBP and CNN with deep learning. The results show that CNN performs better and is more robust among the method of extracting features.