This paper presented an efficient cepstral feature extraction
method for pattern recognition applications. In the proposed
method, images are converted to one dimensional signals and
the MFCCs and their corresponding polynomial coefficients are extracted from them. A database of the cepstral features of pattern
images is produced in the training phase and used for feature
matching in the testing phase after interpolation with different
types of interpolation methods. Experimental results reveal
that the proposed technique, which is comonly used for speaker
identification, can be employed for feature extraction from one
dimensional images. Feature extraction from the different transform
domains has been tested, and it was noticed that extracted
features from the DCTs of the one dimensional pattern signals are
the most powerful among all other features. This is due to the compaction
of energy inherent in the DCT, which produces features
from the 1st few samples after DCT. The results have shown that
the recognition probability can increase to 100% for patterns in the
noise-free environment.