Hierarchical sparse representation for object recognition

Abstract

Recently, generic object recognition that achieves human-like vision has being looked to for use in robot vision, automatic categorization of images, and image retrieval. In object recognition, semi-supervised learning, which incorporates a large amount of unsupervised training data (unlabeled data) along with a small amount of supervised data (labeled data), is regarded as an effective tool to reduce the burden of manual annotation. However, some unlabeled data in semi-supervised models contain outliers that negatively affect the parameter estimation during the training stage. Such outliers often cause an over-fitting problem especially when a small amount of training data is used. Furthermore, another problem that occurs when using the conventional methods is that when labeling an image based on super-pixel representation, the lack of discrimination of the image features and the scale variance of the objects decreases the recognition accuracy because the feature extraction is based on the mono-scale segmentation. In this paper, we propose an object recognition method for solving both problems. For the former problem, our method prevents the over-fitting associated with the semi-supervised based approach by using sparse representation to suppress existing outliers in the data. For the latter problem, we employ Tree Conditional Random Field to construct the hierarchical structure of an image. Experiment results using two datasets confirm the effectiveness of our method.