基于关联图关键边发现的人脸图像聚类算法

针对真实场景中大量类别数未知、样本数量不均衡、数据分布复杂等导致人脸图像智能提取准确率低的问题，提出了基于关联图关键边发现的人脸图像聚类算法。首先，通过融合多个卷积神经网络提取的图像样本特征，获得鉴别性更强的特征向量，并计算不同样本之间的相似度；然后，利用拒真率和认假率设置合适的门限值，将得到的相似度结果与门限值进行比较，筛选出相似程度高的样本对，并添加样本对之间的连接边来构建关联图；再利用介数中心性测度，设计关键边发现方法，挖掘关联图中可能连接不同簇的重要连接边；最后，采用图卷积网络确认是否存在上述重要连接边以获得最终的聚类簇。实验结果表明，所提算法能够提高人脸图像聚类的准确率。

Face Clustering Algorithm Based on Key Links Discoveryof Affinity Graph

In real scenes, unknown number of categories, unbalanced number of samples and complexdata distribution can lead to low accuracy of intelligent face image extraction. To solve these problems, aface image clustering algorithm based on the discovery of key edges of association graph is proposed.First, the algorithm merges features of image samples extracted by multiple convolution neural networks toobtain new feature vectors with stronger capability of identification, and then calculates the similaritybetween different samples. Then, an appropriate threshold value is set for the rejection rate and therecognition rate, and the similarity result obtained in the previous stage is compared with the thresholdvalue. The sample pairs with high similarity degree are screened out, and the association graph isconstructed by adding the connecting edge between the above sample pairs. Using the intermediatenumber centrality measure, the key edge discovery method is designed to dig the important connectingedges that may connect different clusters in the association graph. Finally, the graph convolution networkis used to confirm the existence of the above important connection edges, such that the final cluster isobtained. Experimental results show that the proposed algorithm can improve the accuracy of face imageclustering.