Human Object Classification Using Dual Tree Complex Wavelet Transform and Zernike Moment - Transactions on Large-Scale Data-and Knowledge-Centered Systems - page 95

method, we have also experimented with multilevel discrete wavelet transform coeffi-
cients as a feature set, for human object classification, and confusion matrix is given
in Table 4. We have also evaluated the confusion matrix for other state-of-the-art
methods proposed by Dalal and Triggs [7], Lu and Zheng [9], Renno et al. [14] and
Chen et al. [15], and confusion matrix for them are given in Table 5.

Table 1. Confusion matrix for the proposed method using Zernike moment as a feature set

Zernike moment as a feature set

Predicted classes

True classes

Human Non-human

Human 81 19

Non-human 27 73

Table 2. Confusion matrix for the proposed method using Dual tree complex wavelet transform
(DTCWT) coefficients at multilevel as a feature set [Khare et al. [2]]

DTCWT coefficients (Level - 1) as
a feature set

DTCWT coefficients (Level - 2) as
a feature set

Predicted classes Predicted classes

True classes True Classes

Human Non-human Human Non-human

Human 90 10 Human 92 8

Non-human 10 90 Non-human 10 90

DTCWT coefficients (Level - 3) as
a feature set

DTCWT coefficients (Level - 4) as
a feature set

Predicted classes Predicted classes

True classes True classes

Human Non-human Human Non-human

Human 95 5 Human 95 5

Non-human 6 94 Non-human 6 94

DTCWT coefficients (Level - 5) as
a feature set

DTCWT coefficients (Level - 6) as
a feature set

Predicted classes Predicted classes

True classes True classes

Human Non-human Human Non-human

Human 97 3 Human 99 1

Non-human 6 94 Non-human 3 97

DTCWT coefficients (Level - 7) as a feature set
True classes

Predicted classes

Human Non-human

Human 99 1

Non-human 1 99

From Table 3, we found that the proposed method gives better human object
classification result at higher levels of Dual tree complex wavelet transform
coefficients after combining Zernike moment instead of only Dual tree complex
wavelet transform coefficients as shown in Table 2 (Khare et al. [2] method results).
From Table 4, we found that by using only discrete wavelet transform as feature,
human objects have more confusion with non-human objects as compared with the
results obtained by the proposed method. Confusion matrix result for single feature
set as given in Table 1, 2 and 4 gives poor results in comparison with combination of