Fingerprint-Based Gender Classification

of 6

Please download to get full document.

View again

All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
PDF
6 pages
0 downs
0 views
Share
Description
Fingerprint-Based Gender Classification
Tags
Transcript
  Fingerprint-Based Gender Classification Ahmed Badawi 1 , Mohamed Mahfouz 1 , Rimon Tadross 1 , Richard Jantz 2 1 Biomedical Engineering Department, University of Tennessee Knoxville 2 Anthropology Department, University of Tennessee Knoxville    Abstract - Gender classification from fingerprints is an important step in forensic anthropology in order to identify the gender of a criminal and minimize the list of suspects search. A dataset of 10-fingerprint images for 2200 persons of different ages and gender (1100 males and 1100 females) was analyzed. Features extracted were; ridge count, ridge thickness to valley thickness ratio (RTVTR), white lines count, ridge count asymmetry, and pattern type concordance. Fuzzy C- Means (FCM), Linear Discriminant Analysis (LDA), and Neural Network (NN) were used for the classification using the most dominant features. We obtained results of 80.39%, 86.5%, and 88.5% using FCM, LDA, and NN, respectively. Results of this analysis make this method a prime candidate to utilize in forensic anthropology for gender classification in order to minimize the suspects search list by getting a likelihood value for the criminal gender  . Keywords :  Gender classification, Fingerprint, White Lines, Forensic Anthropology. 1.0 Introduction   Fingerprint identification and classification has been extensively researched in the literature [1], however very few researchers have studied the fingerprint gender classification problem [2-13]. Acree used the ridge density, defined as the number of ridges in a certain space; it was shown that the females have higher ridge density [2]. Kralik showed that the males have higher ridge breadth, defined as the distance between the centers of two adjacent valleys, than females [3]. Two studies showed that the males have higher ridge count than the females [4-5]. It was shown that both males and females have higher rightward directional asymmetry in the ridge count [5-8], with the asymmetry being higher in males than females [8], and higher incidence of leftward asymmetry in females [5]. Female's fingerprints are significantly of lower quality than male fingerprints [9]. The appearance of white lines and scars in fingerprint images is very common in housewives [10]. In this paper we studied the different debates in the literature for the few articles that exist concerning the significance of ridge count, pattern type concordance, ridge count asymmetry, ridge thickness to valley thickness ratio (RTVTR), and white lines count features on the classification performance. We analyzed different features that can be significant in gender classification and different classifiers performances. 2.0 Materials and Methods In our gender classification analysis from fingerprints, we acquired the data first then we extracted the whole features for every finger, averaged the features for the  person’s 10 fingers, and classified the gender of each  person based on different combinations of these features. The overall features include ridge count, RTVTR, fingerprint pattern type, white lines count,  pattern type concordance between the corresponding left-right fingerprints, and ridge count asymmetry  between the left-right corresponding fingerprints. Statistical analysis was performed for pattern types, ridge count, and ridge counts along pattern types. 2.1 Dataset   A dataset of 10-fingerprints for 2200 persons of different ages and gender (1100 males, and 1100 females) were scanned from their ink prints as shown in figures 1 and 2, and were analyzed for the ridge count, and pattern type features. The RTVTR, and white lines count features were analyzed for 255 persons (150 males, and 105 females). 2.2 Features Extraction   2.2.1 Ridge and Valley Thicknesses The average ratio between the ridge thickness and the valley thickness for each of the subject’s fingerprints was calculated automatically, and an average ratio was calculated for every subject. The fingerprint image was divided into 30x30 non overlapping blocks. The local ridge orientation within each block was calculated [16] as shown in Figure 3. The projection profile of the valleys and ridges along a line perpendicular to the local ridge orientation in each block was calculated, and the  projection profile was binarized using 1D optimal thresholding [17]. The resultant binary profile represents the ridges and valleys in this block, the high  binary value represents the valleys and the low binary value represents the ridges. Figure 4 shows two blocks of female and male fingerprints, and their projection and  binary profiles. The average RTVTR was calculated for each block. The uniformity of ridges and valleys within the blocks varies, for blocks having non uniform ridges and valleys due to the low quality of the fingerprint image in this region, the ridge orientation estimation is   Figure1: Two different fingerprints for a male showing no (or few) white lines and small RTVTR usually incorrectly estimated, and thus the RTVTR calculated for this block is incorrect, so only the blocks having the best quality should contribute to the average RTVTR calculated for this fingerprint. For each block, a quality index was calculated as the average difference  between the values of successive singular points (Minimas and Maximas) of the projection profile,  blocks of good quality have higher quality index than those of bad quality. Figure 5 shows a good quality  block, having quality index of 0.244, and correctly estimated RTVTR of 1.67, and a bad quality block, having quality index of 0.061, and incorrectly estimated RTVTR of 1.06. The blocks were arranged in a descending order based on their quality index, and the RTVTR of the best 15 were averaged and taken as the average RTVTR for this fingerprint. 2.2.2 White Lines Count, Ridge Count, Pattern Type, Pattern Type Concordance, and Ridge Count Asymmetry   The white lines count and ridge count were extracted manually for each fingerprint, an average white lines count as well as the ridge count was calculated for each subject. Pattern type was extracted manually for each fingerprint, and the pattern type concordance was calculated for the fingerprints of each right-left corresponding fingerprint pair for the subject, such that the concordance value is 1 if the corresponding Figure 2: Two different fingerprints for a female showing large count of white lines and large RTVTR. (a) (b) Figure 3: (a) a male fingerprint, and (b) a female fingerprints.  fingerprints have the same pattern type, and is 0 otherwise, then the sum of the 5 fingerprint pairs concordance values was calculated. The ridge count asymmetry between the right-left corresponding fingerprints for a subject was calculated, the asymmetry is 1 for a left-right corresponding fingerprint pair if the ridge count of the left fingerprint is greater than the right one, is –1 if it is smaller, and is 0 if both ridge counts are equal. The sum of the asymmetry values of the 5 fingerprint pairs of the subject was calculated. (a) (b) Figure 4: Block from (a) a male’s fingerprint with RTVTR of 0.54, and (b) a female’s fingerprint with RTVTR of 2.33. (a)   (b) Figure 5: (a) Good, and (b) bad quality blocks, and their profiles. 3.0 Results 3.1 White Lines Count and RTVTR Statistics The female’s fingerprint is characterized by a high RTVTR, while the male’s fingerprint is characterized  by low RTVTR, with the exception of small percentage of male’s fingerprints having high RTVTR, and female’s fingerprints having low RTVTR. Figure 6 shows histograms of the RTVTR of the females, with µ=1.05473, σ =0.1245, and the males, with µ=0.9494, σ =0.1045, with t-value=6.866, and ρ -value=5.158e-11. The following histograms have a horizontal axis of the feature values and the vertical axis is the number of cases sharing the same feature value. (a)   (b) Figure 6: Histograms of the RTVTR for (a) females, and (b) males The female’s fingerprint is characterized by high count of white lines, with the exception of small percentage having few or no white lines. The male’s fingerprint is characterized by having no or few number of white lines, with the exception of small percentage having high count of white lines. Figure 7 shows histograms of the white lines count for the females, with µ=   6.4721, σ =4.6139, and the males, with µ=1.1617, σ =1.5869, with t-value=13.27, and ρ -value=2.387e-30 (a)   (b) Figure 7: Histograms of white lines count for (a) females, and (b) males. 3.2 Ridge Count, Ridge Count Asymmetry, and Pattern Type Concordance Statistics The male’s ridge count is slightly higher than the female’s, with high standard deviation in its distribution among both genders. Figure 8 shows histograms of the ridge count for the females, with µ=13.6671, σ =4.9845,  and the males, with µ=14.6914, σ =4.9336, with t-value =4.802, and ρ -value=1.685e-06. (a) (b) Figure 8: Histograms of the ridge count for (a) the females, and (b) the males. The ridge count asymmetry between left and right hand fingerprints shows high variance value for both females and males, and slight difference in the mean value  between females and males, with the females having slightly higher degree of rightward asymmetry. Figure 9 shows histograms of the asymmetry values for the females, with µ=-0.5333, σ =2.2148, and the males with µ=-0.8333, σ =2.2056, with t-value=3.155, and ρ -value=0.00163. (a) (b) Figure 9: Histograms of the ridge count asymmetry for (a) the female, and (b) the males. The pattern type concordance doesn’t show significant variation among males and females fingerprints. Figure 10 shows histograms of the concordance values for the females, with =2.8991, σ =1.1049, and the males with µ=2.8104, σ =1.09, with t-value =1.879, and ρ -value=0.06041. (a) (b) Figure 10: Histograms of the pattern type concordance for (a) the female, and (b) the males. 3.3 Pattern Type Statistics   The overall number of fingerprints belonging to the different 25 fingerprint types shown in figures 11-12 was analyzed; the Ulnar loop type is the most abundant type, followed by the Monocentric whorl.   The  probability of occurrence for each pattern type is nearly the same for males and females with slight differences, as shown in figure 11. Figure 11: Male/Female number of fingerprints per  pattern type. The ridge count varies among the pattern types. The mean ridge count for each pattern type varies slightly  between males and females, as shown in figure 12. Figure 12: Male/Female mean ridge count. 3.4 Classification First we applied Fuzzy C-Means as an unsupervised clustering method for overall feature vectors, then we applied linear discriminant analysis on the data (LDA) and finally we applied Neural Network classifier on different combinations of the extracted features. We found that the most significant features are: the white lines count, RTVTR. The ridge count has shown to have little significance, and slightly enhance the classification with the Neural Network classifier, while it degrades the  performance of the Fuzzy C-Mean classifier. Adding any of the other least significant features to the input vectors result in degradation of performance of the classifiers.  3.4.1 Fuzzy C-Means (FCM) Clustering By applying the FCM algorithm on the white lines count and RTVTR only as input features, a result of 80.39% classification rate was obtained and this result is shown as a confusion matrix in table 1. Adding ridge count to the previous two features, we obtained a degraded result of 56.47% classification as shown in table 2. It is shown in table 2 that adding ridge count to the feature vector highly decreased the classification rate of males and slightly decreased the females’ classification rate. Table 1: Confusion matrix for the FCM classification  based on the white lines count, and RTVTR features. Actual  \  Estimated   Males Females Total Males 145 5 150 Females 45 60 105 Total 190 65 255 Table 2: Confusion matrix for the FCM classification  based on the white lines count, RTVTR, and ridge count features. Actual  \  Estimated   Males Females Total Males 88 62 150 Females 49 56 105 Total 137 118 255 3.4.2 Linear Discriminant Analysis (LDA) By applying the Linear Discriminant Analysis on the white lines count and RTVTR, we got a result of 84.3 % classification rate for the testing set, as shown in the confusion matrix in table 3, with training error rate of 0.18. By adding the ridge count to the previous features, we got a result of 86.5 %, as shown in table 4, with error rate of 0.18. Table 3: Confusion matrix for the testing set for the LDA based on white lines count, and RTVTR.   Actual  \  Estimated   Males Females Total Males 50 2 52 Females 12 25 37 Total 62 27 89 Table 4: Confusion matrix for the testing set for the LDA based on white lines count, RTVTR, and ridge count   Actual  \  Estimated   Males Females Total Males 50 2 52 Females 10 27 37 Total 60 29 89 3.4.3 Neural Network Classification A Multi Layer Back Propagation Neural Network as a non-linear classifier was used [13] for this gender classification analysis. We divided our data to 2/3 for training and 1/3 for testing and we studied the effect of different combinations of the extracted features on the classification rate. The best number of hidden neurons for each features combination was determined empirically. We used the RTVTR, and the white lines count features as the inputs for the Neural Network with 5 hidden layer neurons, learning rate of 0.2, and momentum term of 0.4. The classification rate for the training set was 88.55% , and for the testing set was   86.5% ,  with training root mean square error (RMS) of 0.09 in 3000 epochs. Results of the training and testing confusion matrices are shown in tables 5, and 6 respectively. Table 5: Confusion matrix for training set for the Neural  Network classification based on the white lines count, and RTVTR features. Actual  \  EstimatedMales Females Total Males 92 6 98 Females 13 55 68 Total 105 61 166 Table 6: Confusion matrix for the testing set for the  Neural Network classification based on the white lines count, and RTVTR features. Actual  \  EstimatedMales Females Total Males 47 5 52 Females 7 30 37 Total 54 35 89 We found that using the RTVTR, the white lines count, and the ridge count features as input to the Neural  Network, slightly improves the performance giving; 89.16% for the training set, and 87.64% for the testing set, with training root mean square error of 0.09 in 2000 epochs, learning rate of 0.2, momentum term of 0.5, hidden layer of 9 neurons. The training and testing confusion matrices are shown in tables 7, and 8 respectively. Table 7: Confusion matrix for training set for the Neural  Network classification based on the white lines count, RTVTR, and ridge count features. Actual  \  EstimatedMales Females Total Males 91 7 98 Females 11 57 68 Total 102 64 166 Table 8: Confusion matrix for the testing set for the  Neural Network classification based on the white lines count, RTVTR, and ridge count features. Actual  \  EstimatedMales Females Total Males 47 5 52 Females 6 31 37 Total 53 26 89
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks