基于改进Adaboost集成学习的空间目标识别

针对空间目标的不合作性特点以及Adaboost集成学习算法的过拟合问题, 提出了一种基于组合特征和改进Adaboost的空间目标图像识别算法. 将空间目标图像的几何特征和变换特征进行融合, 从不同的方面更精确地描述目标信息, 并对Adaboost算法进行改进, 根据样本在权重上的分布情况, 在训练时进行分段更新权重, 从而缓解分类器的过拟合现象, 提高目标识别的稳定性. 通过仿真实验证明, 与传统的Adaboost算法相比, 本文算法在空间目标图像识别中取得了更好的效果.

Space Target Recognition Method Based on Improved Adaboost Algorithm

Due to the non-cooperative character of space target and the overfitting of adaboost algorithm under high noises, an space target recognition method based on combined features and improved adaboost is proposed. The combined features which consist of the geometric features and transform features are extracted to describe target information precisely from different aspects. Furthermore, an improved adaboost algorithm is presented, which adopts a new weights updating method piecewisely in the light of the weights distribution of samples. Thus the proposed method can avoid the overfitting problem and improve the robustness of classification. Experiments on space target images showed that the proposed method has better classification capability and obtains higher classification accuracy.