Semantic ear feature reduction for source camera identification

摘要

Energy based Bior 4.4 feature is proven suitable for identifying source camera of ear bio-metric images when a small number of distinct camera sources are used. This level-2 Bior 4.4 feature vector bears 36 energy values. In this paper, a semantic way of reducing this feature vector is discussed which is capable of identifying the source camera of ear biomet-ric images. We analyze the consequences of the reduction towards performance in terms of accuracy. Based on the mean of variances of wavelet energy feature, the size of the feature vector is gradually reduced. Reduction of accuracy of source camera identification is expected with reduction of the feature vector size. However interestingly, we can remove less important feature dimensions without affecting the accuracy much. We need to ensure preserving the feature indices that are deciding factors in yielding the accuracy. From the experiment on 3-class source camera classification, it has been found that even the feature size can be reduced to l/3rd (i.e. up to 12 values from 36 values) with a tolerance of only 1 % degradation in accuracy. Hence we grossly conclude that very low dimensional feature can be potent to predict source camera blindly with good accuracy.