Comparison of multiclass SVM classification techniques in an audio surveillance application under mismatched conditions

In this paper, we compare the performance of classification techniques for multiclass support vector machines in an unstructured environment. In particular, we consider the following methods: one-against-all, one-against-one, decision directed acyclic graph, and adaptive directed acyclic graph. The performance is compared in terms of classification accuracy, training time, and evaluation time. An audio surveillance application is looked at under different noise conditions and varying signal-to-noise ratio with mel-frequency cepstral coefficients and other commonly used time and frequency domain features. The results show that while there isn't much difference in the classification accuracy using the four approaches under clean and low noise conditions, the one-against-all method was found to give relatively better classification accuracy in high noise conditions when trained with clean samples only. However, the results were much more even with multi-conditional training. Also, the training time for the one-against-all approach was found to increase significantly as the training data increased fourfold while the one-against-one approach showed a significantly higher evaluation time.