Autoencoder Based Dimensionality Reduction of Feature Vectors for Object Recognition

Abstract

Object recognition can be performed with high accuracy thanks to the robust feature descriptors defining the significant areas in images. However, these features suffer from high dimensional structure, in other words "curse of dimensionality" for further processes. Autoencoders (AE) are proposed in this study to solve the dimensionality reduction problem of visual features. To assess the efficacy, object recognition is performed using reduced dimensional visual features. For this purpose, dimensionalities of three well-known feature vectors, namely, HOG, SIFT and SURF, are reduced to half. Moreover, deep learning based features are also reduced. Then, reduced vectors, which are called as AE-HOG, AE-SIFT, AE-SURF and AE-DEEP are fed to object recognition task. Also, dimensionality reduction is implemented by a variant of AE, variational autoencoder (VAE) and PCA, which is the most studied unsupervised method for these features, and the results are compared. Furthermore, all experiments are repeated on noisy images. Results suggest that dimensionality reduction of these feature vectors can be accomplished successfully owing to the proposed method.