Evolutionary multi-objective optimization based ensemble autoencoders for image outlier detection

doi:10.1016/j.neucom.2018.05.012

Return

Evolutionary multi-objective optimization based ensemble autoencoders for image outlier detection

2018-10-01

DOI：10.1016/j.neucom.2018.05.012

Original

Save

View PDF

Abstract

En 中文

Image outlier detection has been an important research issue for many computer vision tasks. However, most existing outlier detection methods fail in the high-dimensional image datasets. In order to address this problem, we propose a novel image outlier detection method by combining autoencoder with Adaboost (ADAE). By ensembling many weak autoencoders, our method can better capture the statistical correlations among the features of normal data than the single autoencoder. Therefore, the proposed ADAE is able to determine the outliers efficiently. In order to reduce the many parameters in ADAE, we introduce the Sparse Group Lasso (SGL) constraint into the learning objective of ADAE. We combine Adagrad with Proximal Gradient Descent to optimize this additional learning objective. We also propose the multi-objective evolutionary algorithm to determine the best penalty factors of SGL. By evaluating on several famous image datasets, the detection results testify to the outstanding outlier detection performance of ADAE. The evaluation results also show SGL can make the detection model more compact while maintaining the similar detection performance. (C) 2018 Elsevier B.V. All rights reserved.

Keywords:

Image outlier detection

Autoencoder

Adaboost

Sparse Group Lasso (SGL)

Mutli-objective evolutionary algorithm

Adagrad Proximal Gradient Descent (Ada-PGD)

AI Summary

Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.