【摘要】 提出了一种基于PEDCC-Loss和聚类的方法来提升CNN分类器分类性能的算法。利用CNN以及PEDCCLoss来对图像进行特征提取,然后用BIRCH聚类算法对每类图像的隐特征进行聚类,以获得更好、更逼真的非线性边界,最大程度地减少误分类的边界点。将网络最后一层的PEDCC权重作为每类图像的中心,并以聚类后的子簇的簇心作为分类的辅助判断依据进行图像分类。实验结果表明,该算法的分类准确率相比CNN有一定的提升。更多还原

【Abstract】 This paper proposes an algorithm based on PEDCC-Loss and clustering to improve the classification performance of CNN classifiers.The features are extracted by convolutional neural network and PEDCC-Loss loss function.PEDCC(Predefined Evenly-Distributed Class Centroids) artificially specifies multiple evenly distributed class centroids,which can reach more compact intra-classes distance and more discrete inter-class distance.Then the latent features of each class of image are separately clustered by BIRCH(Balanced Iterative Reducing and Clustering using Hierarchies) clustering algorithm for obtaining a better and more realistic nonlinear boundary to minimize the boundary points of misclassification.The PEDCC weight of the last layer of the network is taken as the center of each type of image,and the centers of the clustered subclusters are used for the auxiliary judgment of classification.更多还原