摘要
为使AP算法对图像进行聚类时充分考虑不同尺度的特征及有效利用未标记数据的特征,提出了结合特征金字塔网络的半监督AP聚类算法(Semi-supervised AP clustering Based on Feature Pyramid Networks, FPNSAP)。FPNSAP算法使用改进的特征金字塔网络来获得图像不同尺度的特征图,对不同大小的特征图进行融合,获得图像的高级语义特征,识别不同大小、不同实例的目标;k近邻标记更新策略可以动态增加标记数据集样本数量,充分利用未标记数据的特征,提高AP算法的聚类性能。FPNSAP算法与四个经典算法(FCH、SAP、DCN和DFCM)在Fashion-MNIST、YaleB和CIFAR-10数据集上进行实验对比,结果表明,FPNSAP算法具有较高的聚类性能,同时算法的鲁棒性更好。
In order to make the AP algorithm fully consider the features of different scales and effectively use the features of unlabeled data when clustering images, a semi-supervised AP clustering based on feature pyramid network (FPNSAP) is proposed. FPNSAP algorithm uses an improved feature pyr-amid network to obtain feature maps of different scales, fuses feature maps of different sizes to ob-tain high-level semantic features of images and identify targets of different sizes and instances;The k-nearest neighbor label updating strategy can dynamically increase the number of labeled data sets, make full use of the characteristics of unlabeled data, and thus improve the clustering perfor-mance of AP algorithm. FPNSAP is compared with four classical algorithms (FCH, SAP, DCN and DFCM) on the Fashion-MNIST, YaleB and CIFAR-10 datasets. The experimental results show that FPNSAP has higher clustering performance and better robustness.
出处
《应用数学进展》
2023年第3期969-979,共11页
Advances in Applied Mathematics
