基于鲁棒鉴别特征学习的图像特征提取算法

Robust discriminant feature learning algorithm for image feature extraction

为了增强图像特征鉴别力和鲁棒性,获取图像紧致特征表达是关键。现有的图像特征学习方法大多采用最大化L_(2)范数的方式定义,导致其对噪声和异常值十分敏感。针对这个问题,提出了一种基于L_(2,1)范数的鲁棒鉴别特征学习算法。该算法在数据预处理中加入了类内聚拢操作,使得同类样本尽可能靠近,减小了类内异常样本以及强噪声样本的影响;此外,引入L_(2,1)范数重新定义了数据的类内和类间相关矩阵,使得模型更具鲁棒性,且提取的特征鉴别能力更强。实验结果显示,相比于现有的一些最新算法,提出的算法不仅具有较高分类准确率,同时还具有较快的收敛速率。这表明了提出算法所提取特征的图像特征具有较强的鉴别力和鲁棒性。

To enhance the discrimination and robustness of image features,the key is to obtain the compact feature expression of the image.However,most of the existing distinguishing feature learning methods were defined by maximizing L_(2) norm,which leads to abnormal sensitivity to noise and outliers.To overcome this problem,a robust discriminate feature learning model defined based on maximizing L_(2,1)norm is proposed.The proposed algorithm adds an intra-class gathering operations in the data preprocessing to make similar samples as close as possible,effectively reducing the influence of abnormal and strong noise samples in the class.Additionally,the introduction of the L_(2,1) norm redefines the intra-class and inter-class correlation matrix of the data,making the model more robust and the extracted features have stronger discriminative ability.The results show that compared with the state-of-the-art algorithms,the proposed algorithm has a higher classification accuracy and simultaneously has a faster convergence rate.This shows that the image features extracted by the proposed algorithm have strong discrimination and robustness.