A novel codebook representation method and encoding strategy for bag-of-words based acoustic event classification

The bag-of-words (BoW) model has been widely used for acoustic event classification (AEC). The performance of the BoW based AEC model is much influenced by "codebook construction" and "histogram generation". The common approaches for constructing the codebook and generating the histogram are the K-means and vector quantization encoding (VQE) respectively. However, they have some inherent disadvantages which pose negative effects on the AEC performance. In this paper, for the BoW based AEC problem, we propose a novel method to construct the codebook and generate the histogram. The self-organizing feature map (SOFM) network is utilized for codebook construction, which can ameliorate the local optimization problem. In addition, an N-Competition encoding strategy is proposed for histogram generation, and the robustness to the boundary points is improved. Experimental result shows that, the proposed method can achieve average 2.4% improvement in accuracy over the traditional BoW based method. Experimental analysis denote that our proposed approach can obtain robust boundary points and effective codebook.