Abstract
In storing large databases of images such as fingerprint and medical databases, the required memory size becomes a great challenge. This work demonstrates a framework for reducing the size of large image databases used in pattern recognition applications with decimation, and reconstructing the images with their original sizes using interpolation for feature extraction. For pattern recognition applications, a new trend based on Mel-Frequency Cepstral Coefficients (MFCCs) is presented in the paper. To reconstruct the images to their original sizes, interpolation methods like bilinear, bicubic, warped-distance, and neural methods are investigated and compared. The sensitivity of the extracted features from the images to the interpolation method used is studied. For the feature extraction process, the interpolated images are converted into one dimensional signals with lexicographic ordering and employed in time domain or transformed to Discrete Wavelet Transform (DWT), Discrete Sine Transform (DST), or Discrete Cosine Transform (DCT) domain. The MFCCs and polynomial shape coefficients are then extracted to generate the database of features, which can be used for pattern identification using neural networks. The pattern recognition is conducted by getting features from the pattern image under test. Experimental results show that feature extraction from an interpolated image to retain the original image dimensions can be used robustly for pattern recognition. In addition, the results reveal that the best domain for feature extraction is the DCT.