부산대학교 도서관

Synthetic aperture radar (SAR) image classification usually relies on sample-based training. Mid-level features are widely used because SAR samples are scarce and objects exhibit complex visual appearances and strong variability. The common mid-level feature model is unsupervised, and the classification results are not robust. In this paper, we introduce mid-level features referred to as attributes. We propose a label-based attributes (L-Attribute) learning model that is based on solving mixed constrained optimization problems. We build discriminative image representations by learning a multi-kernel mapping between the input image features and the attribute space. To solve feature variability, we propose a label-based feature aggregation method to ensure that the attribute representations of the same class are similar and that different classes are far apart. In addition, we enforce that these learned attributes are highly discriminative and easy to predict. The framework is very effective and well generalized in scarce datasets.