Conclusions and Recommendations


As seen from the results, we determined that correlation matching alone is not enough for an accurate identification system. We need some way of limiting the number of correlation calculations as well as a way of making these calculations more accurate. Accuracy and speed are critical in the real world to being competitive in any industry. This is why we wanted to implement a coarse level identification system to classify our database fingerprints into groups. Then, we just classify the input fingerprint and try to match it with only the database fingerprints of their classification. Our coarse level identification system seems to work as evidenced by its performance on the clear input sample. From this result and the fact that the smudged fingerprint gave inconclusive results, we determined that clear input fingerprints are absolutely critical. Unfortunately, finding full size clear fingerprint samples from other sources was very difficult. For our sampled fingerprints, the variations in the amount of ink on each part of the finger obviously contribute error.

If the ink is darker in one place compared to another, the core-delta points can not be extracted. One solution is to implement a feature extraction algorithm, which extracts the lines and ridges from the fingerprint before attempting to classify the fingerprint as a certain types. However, even with a feature extraction algorithm, unclear fingerprints will still introduce large errors. In actual fingerprinting systems, the fingerprints are collected with a much drier ink and special paper. Then, core and delta points can be extracted from the input fingerprint and matched with the core and delta points of the database fingerprints. This helps to line up fingerprints to account for rotation and makes the correlation even more accurate since convolution and the fft only account for translation not rotation.

While the correlation results alone are not enough to identify a fingerprint, if used in conjunction with a type classifier and clear fingerprints a correlation index can be extremely useful. So, in the end, our classification method in series with a correlation-based index of similarity appears to be a good technique of matching fingerprints.

 


[Home] [Abstract][Background][Classification] [Correlation] [Results] [Conclusions][Code][References]