Person identification in Ethnic Indian Goans using ear biometrics and neural networks.

This study presents new insights and experimental results for the use of ears as a non-invasive biometric for human identification. To determine the uniqueness of the external ear pattern two methods were employed: The Weighted Scoring System and Pattern Recognition by Neural Networks. A total of 10 external ear features classified into 37 sub-features for both right and left ears of 400 Indians of Goan origin were studied after acquiring standardized side profile digital photographs. These features were then converted to numeric scores by the 'Weighted Scoring System' which were then compared to ascertain the uniqueness of ear pattern in same and different individuals. Apart from this feature-wise comparison, the initially acquired photographs of 800 individual ears were scrutinized and 80 visually similar ear patterns were found. After appropriate pre-processing of five train and five test images of each of these 80 visually similar ear patterns, the images were analyzed by a specially designed software and 360 feature vectors which were the distances from the centroid to the outer edge of the ear were extracted and saved. The feature vectors of train and test images were employed to train and test the Neural Networks. The result revealed that none of the individuals in the study sample had identical weighted scores when both right and left ear scores were considered in combination or when bilateral comparison was made in the same individual. The digital analysis of visually similar ear images by Neural Networks revealed a recognition rate of 94% with an Equal Error Rate at threshold value of 0.225. The inter-individual match score among train images were found to be less than the intra-individual match scores between train and test images or the differences found in former were more than that in the latter. Also, all intra-individual scores were above the system threshold (0.225) hence accepted as match, while all inter-individual scores were below it and hence rejected as a match. An independent t-test applied to the intra- and inter-individual match scores indicated that the two distributions were significantly different (p<0.0001). Thus, this study has been successful in determining the uniqueness of ear pattern for person identification and in designing and testing software for recognition of ear patterns from side profile photographs.