Feature fusion for efficient object classification using deep and shallow learning

Abstract

Bag-of-Features (BoF) approach have been successfully applied to visual object classification tasks. Recently, convolutional neural networks (CNNs) demonstrated excellent performance on object classification problems. In this paper we propose to construct a new feature set by processing CNN activations from convolutional layers fused with the traditional BoF representation for efficient object classification using SVMs. The dimension of convolutional features were reduced using PCA technique and the bag-of-features representation was reduced by tailoring the visual codebook using a statistical codeword selection method, in order to obtain a compact representation of the new feature set which achieves increased classification rate while requiring less storage. The proposed framework, based on the new features, outperforms other state-of-the-art approaches that have been evaluated on benchmark datasets: Xerox7, UIUC Texture, and Caltech-101.