Using line segments to train multi-stream stacked autoencoders for image classification

Abstract

Recently, deep learning paradigm and models derived from them have achieved outstanding success in many fields in computer vision such as object recognition, image classification and image segmentation. In this work, the authors preprocess images into segments and then extract their geometric information as inputs to stacked autoencoders. A multi-stream framework based on the different geometric feature spaces of the segments is implemented to learn deep geometric representations that have more discriminative powers and generative capabilities. In order to assess the robustness and smoothness of the proposed representation, four representative Geometric Feature Sets (GFSs) are investigated. To further verify the effectiveness of the proposed method, we apply those GFSs for the image classification experiments on four challenging datasets. Given a smaller size of depth, the proposed multi-stream method achieves comparable or better results compared to the best performers.