While many deep convolutional neural networks show promising performance in various classification tasks, multiple objects appearing in very different sizes, shapes, and appearances cause difficulty in multi-label classification using conventional neural networks. In this paper, we introduce a dual aggregated network on pyramidal convolutional features for multi-label classification. The proposed method includes both feature- and classifier-level aggregation to learn discriminant multi-scale information of various target objects in the image. First, the feature-level aggregation collects the convolutional activation maps from the multi-scale pyramid network, and then it densely pools them to take localized features of each object. We elaborately design the feature aggregation method so that the responses from the objects with different sizes, aspect ratios, and shapes are properly reflected the aggregated activation map. Unlike conventional methods, this process does not require the region proposal step, which reduces the computational burden significantly. Second, we introduce the classifier level aggregation algorithm for integrating the multi-object classifier modules. To maximize the discrimination power of each class, we train one-vs-all classifiers for individual classes using the class-wise loss function. For each test image, the scores from the class-wise classifiers are aggregated to get the final multi-label classification result. By combining the above feature- and classifier-level aggregation methods, our network can be trained in an end-to-end fashion, which is not possible for the conventional multi-label classification algorithms using region proposals. Extensive evaluations on PASCAL VOC 2007 and PASCAL VOC 2012 demonstrate that the proposed algorithm outperforms the state-of-the-art methods.

尽管许多深度卷积神经网络在各种分类任务中均显示出令人鼓舞的性能，但以常规神经网络在多标签分类中出现困难的是，大小，形状和外观非常不同的多个对象出现了。在本文中，我们介绍了关于金字塔卷积特征的双重聚合网络，用于多标签分类。所提出的方法包括特征级分类器和分类器级聚合，以学习图像中各种目标对象的判别式多尺度信息。首先，特征级聚合从多尺度金字塔网络中收集卷积激活图，然后将其密集地汇集起来，以获取每个对象的局部特征。我们精心设计了特征聚合方法，以使来自不同大小的对象的响应，长宽比和形状正确反映了聚合的激活图。与常规方法不同，此过程不需要区域建议步骤，从而显着减少了计算负担。其次，我们介绍了用于集成多对象分类器模块的分类器级聚合算法。为了最大化每个类别的辨别力，我们使用逐级损失函数训练单个类别的一对多分类器。对于每个测试图像，将来自类别分类器的分数进行汇总，以获得最终的多标签分类结果。通过结合以上特征和分类器级别的聚合方法，我们的网络可以以端到端的方式进行训练，这对于使用区域提议的常规多标签分类算法而言是不可能的。