RI: Small: Hierarchical Feature Learning by Heterogeneous Networks with Application to Face Verification

学习好的特征是解决计算机视觉问题的关键，例如识别人脸和理解场景。许多计算机视觉研究人员通过为大型数据库中的每幅图像提供语义标签来学习特征，将每幅图像的信息量限制在几个比特。其他人则通过识别图像中常见的图案(如线条、斑点和更复杂的形状)来学习特征，但忽略了语义信息。该项目开发了学习图像中常见特征的算法，并使用一种名为异质网络的新型深度神经网络预测图像在不同空间尺度上的语义。开发的算法允许在中间层合并语义信息。开发的算法不仅可以改变特征学习的方式，还可以指示如何将特征学习扩展到数百万图像的海量数据集。研究小组使用NCSA的千万亿级超级计算机Blue Waters和两个大规模(数百万图像)图像数据集解决了人脸验证中的挑战性问题。这一项目的研究是与本科和研究生教育相结合的。该项目的研究成果在计算机视觉和模式识别领域具有广泛的应用前景。研究团队计划在完成后向研究社区发布在该项目中收集的算法和人脸数据集。

Learning good features is a key to computer vision problems such as recognizing human faces, and understanding scenes. Many computer vision researchers learn features by providing a semantic label for each image in a large database, limiting the amount of information per image to a few bits. Others learn features by identifying common patterns found in images such as lines, blobs, and more complicated shapes, but ignoring semantic information. This project develops algorithms to learn features that are common in images and also predict the semantics of images at various spatial scales using a new type of deep neural network called Heterogeneous Networks. The developed algorithms allow the incorporation of semantic information at intermediate layers. The algorithms developed can not only change how features are learned but also indicate how to scale feature learning to giant datasets of millions of images. The research team addresses challenging problems in human face verification using NCSA's petascale supercomputer, Blue Waters, and two large scale (millions of images) image data sets. The research of this projected is integrated with both undergraduate and graduate education. The results obtained from this project are applicable to a wide range of applications in computer vision and pattern recognition. The research team plans to release algorithms and face data sets collected in this project to research communities once they are finished.