A novel biometric approach to generate ROC curve from the Probabilistic Neural Network

Abstract

The aim of this paper is to present a new method to produce a Receiver Operating Characteristic (ROC) curve from a Probabilistic Neural Network (PNN). Traditionally, an ROC curve has been used widely to report the recognition system measurements. Two main problems arise when using the PNN. Firstly, the PNN outputs are always logical (zeros and one); secondly, a PNN is considered as a multi-class classifier, because it usually has more than one output class. To solve these problems, we suggest a new approach to acquire the score values from the PNN, establish the relationship between the ROC parameters for each class and fusing them to generate one main ROC curve. Personal authentication based on the Finger Texture (FT) biometric has been used to collect the ROC parameters, where three feature extraction methods have been implemented and evaluated: Coefficient of Variance (CV) statistics, Gabor filter followed by the CV calculations and Local Binary Pattern (LBP) followed by the CVs. The results show the accuracy of the Equal Error Rates (EERs) recorded for each ROC graph compared with the actual practical values.

Finger Texture Verification Systems Based on multiple spectrum Lighting Sensors with Four Fusion Levels

“…Furthermore, a new feature extraction method named the Image Feature Enhancement (IFE) was adopted. In 2016, Al-Nima et al [30] suggested a novel approach to establish the Receiver Operating Characteristic (ROC) graph from an Artificial Probabilistic Neural Network (APNN) by employing the FT characteristic. Three feature extraction methods were examined: the Local Binary Pattern (LBP) with a statistical calculation method called the Coefficient of Variance (CV), the Gabor filter with the CV and the statistical CV computations alone in order to present the accuracy of the suggested approach.…”

“…(5) It has been cited that using CV calculations can provide effective descriptions for the variances of the featured images. The resulting images have been partitioned into non-overlapping windows with a fixed size of 5 × 5 and the CV value is computed for each window [20,29,30,33]. To calculate the CV values, the following equations have been exploited [34]:…”

“…The FLF was applied for the completed extracted features, represented by the CVs, before the APNN. For the ScLF, the outputs of the summation layer are considered to be the scores of the APNN, because the summation layer holds the real output values according to [30]. For the DLF, two full systems for the finger images of the two determined spectrum lightings are established and then the final decision of the two APNNs are competed to perform the employed fusion rules.…”

“…2400 images of fingers belong to the WHT spectrum and 2400 images of fingers belong to the BLU spectrum. Firstly, 4000 images were used in the learning stage and 800 images were used in the testing stage as exploited in the state-of-the-art studies [20,[29][30][31]. Secondly, we have changed the number of learning and testing samples (by decreasing the number of learning images and increasing the number of testing images) and we have recorded the obtained evaluations.…”

“…Previous FT studies such as [13,30,31] has utilized 1 testing sample and 5 learning samples for each participant in the MSCASIA database. Obviously, one testing sample is too small comparing with the number of learning samples.…”