Face detection remains a challenging problem due to the high variability of scale and occlusion despite the strong representational power of deep convolutional neural networks and their implicit robustness. To handle hard face detection under extreme circumstances especially tiny faces detection, in this paper, we proposed a multiscale Hybrid Pyramid Convolutional Network (HPCNet), which is a one-stage fully convolutional network. Our HPCNet consists of three newly presented modules: firstly, we designed the Hybrid Dilated Convolution (HDC) module to replace the fully connected layers in VGG16, which enlarges receptive field and reduces its loss of local information; secondly, we constructed the Hybrid Feature Pyramid (HFP) to combine semantic information from higher layers together with details from lower layers; and thirdly, to deal with the problem of occlusion and blurring effectively, we introduced Context Information Extractor (CIE) in HPCNet. In addition, we presented an improved Online Hard Example Mining (OHEM) strategy, which can enhance the average precision of face detection by balancing the number of positive and negative samples. Our method has achieved an accuracy of 0.933, 0.924, and 0.848 on the Easy, Medium, and Hard subset of WIDER FACE, respectively, which surpasses most of the advanced algorithms.

中文翻译：

---

混合金字塔卷积网络的多尺度人脸检测

尽管深度卷积神经网络具有强大的表示能力及其隐含的鲁棒性，但由于规模和遮挡的高度可变性，面部检测仍然是一个具有挑战性的问题。为了处理极端情况下的硬脸检测，特别是微小的人脸检测，本文提出了一种多级混合金字塔卷积网络（HPCNet），它是一个单阶段的全卷积网络。我们的HPCNet由三个新介绍的模块组成：首先，我们设计了混合膨胀卷积（HDC）模块来替代VGG16中的完全连接层，从而扩大了接收范围并减少了其本地信息丢失。其次，我们构造了混合特征金字塔（HFP），将来自高层的语义信息与来自较低层的细节结合在一起。第三，为了有效地解决遮挡和模糊的问题，我们在HPCNet中引入了上下文信息提取器（CIE）。此外，我们提出了一种改进的在线硬示例挖掘（OHEM）策略，该策略可以通过平衡正样本和负样本的数量来提高人脸检测的平均精度。我们的方法在WIDER FACE的Easy，Medium和Hard子集上分别达到0.933、0.924和0.848的精度，这超过了大多数高级算法。