DLANet: A manifold-learning-based discriminative feature learning network for scene classification

Abstract

This paper presents Discriminative Locality Alignment Network (DLANet), a novel manifold-learning-based discriminative learnable feature, for wild scene classification. Based on a convolutional structure, DLANet learns the filters of multiple layers by applying DLA and exploits the block-wise histograms of the binary codes of feature maps to generate the local descriptors. A DLA layer maximizes the margin between the inter-class patches and minimizes the distance of the intra-class patches in the local region. In particular, we construct a two-layer DLANet by stacking two DLA layers and a feature layer. It is followed by a popular framework of scene classification, which combines Locality-constrained Linear Coding–Spatial Pyramid Matching (LLC–SPM) and linear Support Vector Machine (SVM). We evaluate DLANet on NYU Depth V1, Scene-15 and MIT Indoor-67. Experiments show that DLANet performs well on depth image. It outperforms the carefully tuned features, including SIFT and is also competitive to the other reported methods.