Abstract:  Target representation is the key process for hyperspectral target detection. Many methods have been proposed for better target representation. Among these methods, multiple instance-based learning method (MIL) is an effective one as it does not require pixel-level semantic labels. However, traditional MIL-based hyperspectral target representation methods usually cause the instance-level data imbalance because of the limited target instances and too many background instances in positive bags, leading to poor performance on hyperspectral target representation. To overcome this problem, a data imbalanced multiple instance learning-based method is proposed in this paper. First, a positive sample set with the most probably positive sample in each package will be constructed; then, new positive samples will be synthetized to increase the proportion of positive samples in the positive packages, balancing the positive and negative samples in positive packages, improving the representational ability. Experiments over real-world hyperspectral dataset validate the effectiveness of the proposed method and the experiment results show that the proposed method can enforce the balance of the positive packages and learn target representation more accurately which improves the target detection performance.

Key words:  hyperspectral, target representation, multiple instance learning, imbalance