Biometric Standards, Performance, and Assurance Laboratory

Department of Industrial Technology

Henry Classifications: Analysis of Image Quality, Minutiae Count, and Matching Performance

This research is being led by Dr. Elliott

Abstract

This paper presents an analysis of the quality of the image, minutiae count, and overall performance of fingerprint images based on the Henry system of fingerprint classification and the finger’s relative location on the presenting hand. To do this, 50 users submitted 3 images from four fingers (index, middle, ring, and little). The National Institute of Standards and Technology (NIST) Fingerprint Image Software, release 2 (NFIS2) was used to analyze image quality and minutiae count. Neurotechnologija Ltd.’s VeriFinger was used to produce receiver operating characteristics (ROCs) in order to analyze performance. Our results show differences not only in image quality and minutiae count, but also matching performance based on Henry system classification and finger location.

One thousand sixty-three images were successfully classified; the remaining 62 images were placed in the “unidentified” group. The 1,063 classified images were categorized as follows:

• Whorl: 342 (32.1%)
• Left loop: 291 (27.3%)
• Right loop: 275 (25.8%)
• Tented arch: 106 (9.9%)
• Arch: 49 (4.6%)

After the images had been classified, the next step was to statistically analyze the images by classification and by finger location. Using the data from NFIS2, analysis of variance (ANOVA) tests were performed for the following scenarios:
Sorted by classification:

• NFIQ score (ANOVA #1)
• NFIS minutiae (ANOVA #2)

Sorted by finger location:

• NFIQ score (ANOVA #3)
• NFIS minutiae (ANOVA #4)

The final step in the process was to generate receiver operating characteristics (ROCs) to compare the performance of the images, sorted by classification and finger location.

Statistical Results

ANOVA #1 examined the statistical relationship between the NFIQ score (quality) and the Henry system. These results indicated a p-value of 0.000 at an α level of 0.05 with an R2 value of 6.83%. The confidence intervals are shown below.

ANOVA #1 confidence intervals

ANOVA #2 examined the minutiae count compared against the Henry classification. The results showed a statistically significant difference across all of the classifications, with a p-value of 0.000 at an α level of 0.05 with an R2 value of 11.65%. The confidence intervals are shown below. 

ANOVA #2 confidence intervals

ANOVA #3 examined the NFIQ score (quality) compared to the finger location. This too showed a statistically significant difference across all of the fingers, with a p-value of 0.000 at an α level of 0.05 with an R2 value of 1.31%. The confidence intervals are shown below. 

ANOVA #3 confidence intervals

ANOVA #4 examined the minutiae count compared to finger location, which also showed a statistically significant difference of minutiae across all of the fingers, with a p-value of 0.000 at an α level of 0.05 with an R2 value of 2.99%. The confidence intervals are shown below. 

ANOVA #4 confidence intervals

While all of the ANOVA models showed significance, the one with the highest R2 values was the comparison of minutiae count against Henry classifications. This showed that minutiae count across the classifications is more statistically significant than image quality across classifications and far more significant than finger location with image quality and minutiae count.

Performance Analysis

While the analysis of variance shows the statistical significance of a response against a factor, ROC curves show the performance of the biometric system (i.e., every fingerprint of each finger is matched with every fingerprint of each other finger within the specific classification). The figure below shows the individual ROC graphs from the Henry classification groups, overlaid against the overall data set:

ROC based on classification

Conclusions

Our results show a statistically significant difference in the both the quality scores and minutiae counts of fingerprint images sorted by classification and finger location. Furthermore, the ROCs indicate that different classifications and finger locations will result in variations of matching performance. Out of the five Henry classifiers, the whorl performs above average (the whorl curve is below the entire dataset curve) for a majority of the graph, and is the highest ranked in image quality and minutiae. This research shows a correlation between good image quality along with high minutiae count and good matching performance and Henry classifier. It does not show a strong correlation with image quality, minutiae count, and matching performance with finger location. 
From a systems implementation standpoint, while not every user of a fingerprint system may possess a whorl classification; using whorl classifications over the others will improve the performance of the system. At the very least, an index finger would be preferable over other fingers to achieve the best possible system performance. From a system development standpoint, utilizing knowledge about how different classifications and finger locations will perform; the system can be fine-tuned accordingly. Adjustments can be made in the form of a weighted algorithm to be more or less stringent during the matching process, based on the images being compared. In looking outside of the fingerprint recognition system, a similar weighted algorithm could be used with other forms of identification, including other biometric modalities.
Combining the results produced by the fingerprint system with another modality, such as iris recognition or face recognition, would be a form of biometric fusion. One of the biggest systems to use multiple biometric modalities at one time is the United States Visitor and Immigrant Status Indicator Technology (US-VISIT); which uses both fingerprint and face recognition at national borders and major ports of entry. The goal of the US-VISIT program is to facilitate legitimate travel and trade while apprehending known criminals and terrorists. Currently, US-VISIT does not use biometric fusion.

Conference Proceedings