Performance evaluation of typical unsupervised feature learning algorithms for visual object recognition

• Published in: Proceeding of the 11th World Congress on Intelligent Control and Automation pp. 5191 - 5196
• Main Authors: Shaohua Zhang, Hua Yang, Zhouping Yin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.06.2014
• Subjects: Accuracy; Classification algorithms; deep learning; Dictionaries; Encoding; evaluation; Filtering; Learning systems; object recognition; Training; unsupervised feature learning
• DOI: 10.1109/WCICA.2014.7053598

Many kinds of feature learning algorithms have been proposed and have continually refreshed the state-of-the-art performance in speech recognition, visual object recognition et al. However, most of them are complicated and hard to train, which limits more widely application. In this paper, we present the usability performance evaluation of six typical unsupervised feature learning algorithms from aspects of accuracy, time cost, and hyper-parameters. A common patch based framework [1] with mediocre parameters is adopted to highlight the difference between algorithms. The experiments confirm that sparse coding can attain consistent performance across different datasets. Moreover, random patches with soft threshold function and K-means combining with triangle coding achieve comparable performance with sparse coding, and even faster and easier to train, the results suggest they are good choices to build an application system in practice.