Signature Recognition And Verification Systems

ABSTRACT

In this paper, we present an off-line signature recognition and verification system which is based on moment invariant method and ANN. Two separate neural networks are designed; one for signature recognition, and another for verification (i.e. for detecting forgery). Both networks use a four-step process. First step is to separate the signature from its background. Second step performs normalization and digitization of the original signature. Moment invariant vectors are obtained in the third step. And the last step implements signature recognition and verification. Signature recognition and verification involves two separate but strongly related tasks: one of them is identification of the signature owner, and the other is the decision about whether the signature is genuine or forged. Also, depending on the need, signature recognition and verification problem is put into two major classes: (i) online signature recognition and verification systems (SRVS) and (ii) offline SRVS.

INTRODUCTION

Signature is a special case of handwriting which includes special characters and flourishes. Many signatures can be unreadable. They are a kind of artistic handwriting objects. However, a signature can be handled as an image, and hence, it can be recognized using computer vision and artificial neural network techniques. Signature recognition and verification involves two separate but strongly related tasks: one of them is identification of the signature owner, and the other is the decision about whether the signature is genuine or forged. Also, depending on the need, signature recognition and verification problem is put into two major classes: (i) online signature recognition and verification systems (SRVS) and (ii) offline SRVS. Online SRVS requires some special peripheral units for measuring hand speed and pressure on the human hand when it creates the signature. On the other hand, almost all off-line SRVS system relies on image processing and feature extraction techniques. In the last two decades, in parallel with the advancement in the sensor technology, some successful online SRVS are developed. There are also many studies in the area of offline SRVS category. These studies are generally based on ANN, analysis of the geometry and topology of the signature, and its statistical properties. In this study, we present an off-line signature recognition and verification system which is based on a moment invariant method. Two neural networks are designed; one for signature recognition, and another for verification   (i.e. for detecting forgery).

ARTIFICIAL NEURAL NETWORKS

The human brain provides proof of the existence of massive neural networks that can succeed at those cognitive, perceptual, and control tasks in which humans are successful. The brain is capable of computationally demanding perceptual acts (e.g. recognition of faces, speech) and control activities (e.g. body movements and body functions). The advantage of the brain is its effective use of massive parallelism, the highly parallel computing structure, and the imprecise information-processing capability. The human brain is a collection of more than 10 billion interconnected neurons. Each neuron is a cell that uses biochemical reactions to receive, process, and transmit information. Treelike networks of nerve fibers called dendrites are connected to the cell body or soma, where the cell nucleus is located. Extending from the cell body is single long fibers called the axon, which eventually branches into strands and sub strands, and are connected to other neurons through synaptic terminals or synapses. The transmission of signals from one neuron to another at synapses is a complex chemical process in which specific transmitter substances are released from the sending end of the junction. The effect is to raise or lower the electrical potential inside the body of the receiving cell. If the potential reaches a threshold, a pulse is sent down the axon and the cell is ‘fired’. Artificial neural networks (ANN) have been developed as generalizations of mathematical models of biological nervous systems. A first wave of interest in neural networks (also known as connectionist models or parallel distributed processing) emerged after the introduction of simplified neurons by McCulloch and Pitts (1943). The basic processing elements of neural networks are called artificial neurons, or simply neurons or nodes. In a simplified mathematical model of the neuron, the effects of the synapses are represented by connection weights that modulate the effect of the associated input signals, and the nonlinear characteristic exhibited by neurons is represented by a transfer function. The neuron impulse is then computed as the weighted sum of the input signals, transformed by the transfer function. The learning capability of an artificial neuron is achieved by adjusting the weights in accordance to the chosen learning algorithm.

IMAGE PROCESSING

            The camera-captured or scanned real world images containing human signatures are processed using several image processing algorithms before the calculation of the moment invariants. These processes are given below.

Converting Color image to gray scale image

            In present technology, almost all image capturing and scanning devices use color. Therefore, we also used a color scanning device to scan signature images. A color image consists of a coordinate matrix and three color matrices. Coordinate matrix contains x, y coordinate values of the image. The color matrices are labeled as red (R), green (G), and blue (B). Techniques presented in this study are based on grey scale images, and therefore, scanned or captured color images are initially converted to grey scale using the following equation:

Gray color = 0.299*Red + 0.5876*Green+0.114*Blue……………….. 

Noise Reduction

            Noise reduction (also called “smoothing” or “noise filtering”) is one of the most important processes in image processing. Images are often corrupted due to positive and negative impulses stemming from decoding errors or noisy channels. An image may also be degraded because of the undesirable effects due to illumination and other objects in the environment. Median filter is widely used for smoothing and restoring images corrupted by noise. It is a non-linear process useful especially in reducing impulsive or salt-and-pepper type noise. In a median filter, a window slides over the image, and for each positioning of the window, the median intensity of the pixels inside it determines the intensity of the pixel located in the middle of the window. Different from linear filters such as the mean filter, median filter has attractive properties for suppressing impulse noise while preserving edges. Median Filter is used in this study due to its edge preserving feature.

Background elimination and border clearing

            Many image processing applications require the differentiation of objects from the image background. Thresholding is the most trivial and easily applicable method for this purpose. It is widely used in image segmentation. Thresholding is choosing a threshold value T and assigning 0 to the pixels with values smaller than or equal to T and 1 to those with values greater than T. We used thresholding technique for differentiating the signature pixels from the background pixels. Clearly, in this application, we are interested in dark objects on a light background, and therefore, a threshold value T, called the brightness threshold, is appropriately chosen and applied to image pixels f(x, y) as in the following:

            Signature image which is located by separating it from complex background image is converted into binary image white background taking the pixel value of 1.Vertical and horizontal (histogram) projections are used for border clearing. For both direction, vertical and horizontal, we counted every row zeros and the resulting histogram is plotted sideways.

Signature normalization

            Signature dimensions may vary due to the irregularities in the image scanning and capturing process. Furthermore, height and width of signatures vary from person to person and, sometimes, even the same person may use different size signatures. First, we need to eliminate the size differences and obtain a standard signature size for all signatures. After this normalization process, all signatures will have the same dimensions. In this study, we used a normalized size of 40x40 pixels for all signatures that will be processed further. During the normalization process, the aspect ratio between width and height of a signature is kept intact. Normalization process made use of the following equations:

In these equations:

 xi, yi         :        pixel coordinates for the normalized signature,

 x’i, y’i       :         pixel coordinates for the original signature,

    M          :         one of the dimensions (width or height) for the

ANN DESIGN FOR SIGNATURE RECOGNITION AND VERIFICATION

            We designed a multilayer feed forward artificial neural network for recognition of off-line digitized signatures. The proposed ANN consists of 14 input variables, 18 hidden neurons, and 30 output variables and it is designed to recognize one signature at a time. Back propagation algorithm is used for training.

Training for signature recognition:

            First, an input/output database is created manually for training and testing the ANN for six signature image which are belong to same person but signed different time. Each input vector consists of seven moment invariants obtained for a signature. As explained earlier in section 4, six different moment invariant vectors are produced for each signature. These six vectors are divided into two sets each containing three vectors. One of these sets (3 input vectors) is used in the training of ANN and the other set (remaining 3 input vectors) is used for testing. Additional, we also produce seven extra properties for a signature. As also explained earlier in section 4. The database contained a total of 30 different signature images which are used for both training and testing. Since 3 input vectors for each image is used for training purposes, there are a total of 90 (30*3) input vectors (data sets) in the training set. The remaining 90 data sets are used for testing. ANN contained 30 binary output values each corresponding to one signature being tested as shown in Table 2. Under normal (correct) operation of ANN, only one output is expected to take a value of “1” indicating the recognition of a signature represented by that particular output. The other output values must remain zero. In general, the number of outputs must be equal to the number signatures being considered. Table 2 shows a number of real input/output vectors used in the training set. Which are obtained from a set of signatures? 

Signature verification

            In this part of the study, our purpose is to authenticate a signature i.e. to verify that the signature is not counterfeit and it really belongs to the person who is claimed to be the owner of the signature. The ANN used for this purpose is also a multilayer feed forward network which consists of 14 input variables, 10 hidden neurons, and 2 output variables indicating whether the signature is fake or true. Back propagation algorithm is used for training. The training data set is obtained from three original (authenticated) signatures provided by the real owner and three fake signatures. As it was done for the preparation of the training

Data for the ANN used in recognition, three invariant vectors per signature is used in the training set. Therefore, a total of 18 moment invariant vectors are used in the training set. A sample set of three signatures belonging to the same person.

IMPLEMENTATION AND TEST RESULTS

Signature recognition

            The program used a windows interface as shown in Figure 5. This software allowed the signature images to be loaded one at a time and used in training and testing. First, the signature image is captured using a CCD camera or a scanner, then, through several image processing operations, it is converted to binary and normalized to a 40*40 image as explained earlier. Moment invariant and additional values are obtained from the normalized image which is then used as the input vector to the ANN. After the training of the ANN for signature recognition, the system is ready to recognize a given signature. 

            The signature recognition system is tested using 30 signatures chosen at random. The images were obtained using the following properties:

Signatures were signed inside a special framed area.

Images were taken with a simple CCD camera and they were shot from a fixed distance.

            As explained in Section 4.1., 30 images in our database belonging to 30 different signatures are used for both training and testing. Since 3 (out of 6) input vectors for each image were used for training purposes, there are only 90 (30*3) input vectors (data sets) left to be used for the test set. Under normal (correct) operation of the ANN, only one output is expected to take a value of “1” indicating the recognition of a signature represented by that particular output. The other output values must remain zero. The output layer used a logic decoder which mapped neuron outputs between 0.5-1 to a binary value of 1. If the real value of an output is less than 0.5, it is represented by a “0” value. The ANN program recognized all of the 30 signatures correctly. This result translates into a 100% recognition rate. We also tested the system with 10 random signatures which are not contained in the original database. Only two of these signatures which are very similar to at least one of the 30 stored images resulted in “false positives” (output > 0.5) while the remaining 8 are recognized 

correctly as not belonging to the original set (the output value was <= 0.5). Since recognition step is always followed by the verification step, these kinds of false positives can be easily caught by our verification system. In other words, the verification step serves as a safeguard against “false positives” as well as “false negatives”.

Testing the verification system

            Training for verification is explained in Section 4.2. Signatures used for testing the verification system are obtained the same way as in the recognition system. We tested the verification software using 10 signatures; 5 imitations (counterfeit signatures) and 5 true signatures. The program detected (classified) 4 true signatures and 5 counterfeits correctly. In other words, all counterfeit signatures are detected correctly. Only one signature is classified as a counterfeit while it was not (i.e. a “false negative”). Obviously, a “false negative” should be more acceptable in comparison to a “false positive”, because the person can always be given a second chance to prove that the signature is his/hers. On the other hand, a false positive in verification carries a lot of risk.

LIMITATION

Moment invariants and some global properties which are used as input features for the NN are obtained in the third step. Two separate ANNs are used; one for signature recognition and another for verification. Our recognition system exhibited a 100% success rate by identifying correctly all of the 30 signatures that it was trained for. However, it exhibited poor performance when it recognize Load signature Train Test Recog. ANN T.5 was presented with signatures that it was not trained for earlier. We did not consider this as a “high risk” case, because recognition step is always followed by the verification step and these kinds of false positives can be easily caught by the verification system. Indeed, the verification system did not miss any of the counterfeit Signatures. However, its verification for true signatures lacked some accuracy. We think that this is also acceptable because a person can always be given a second chance to prove the ownership of a signature.

CONCLUSION

In this study, we presented off-line signature recognition and verification system which is based on images processing, moment invariants, some global properties and ANNs. Both systems used a four-step process. In the first step, the signature is separated from its image background. Second step performs normalization and digitization of the original signature. Moment invariants and some global properties which are used as input features for the NN are obtained in the third step. Two separate ANNs are used; one for signature recognition and another for verification. Our recognition system exhibited a 100% success rate by identifying correctly all of the 30 signatures that it was trained for. However, it exhibited poor performance when it recognize Load signature Train Test Recog. ANN T.5 was presented with signatures that it was not trained for earlier. We did not consider this as a “high risk” case, because recognition step is always followed by the verification step and these kinds of false positives can be easily caught by the verification system. Indeed, the verification system did not miss any of the counterfeit signatures. However, its verification for true signatures lacked some accuracy. We think that this is also acceptable because a person can always be given a second chance to prove the ownership of a signature.

Generally, the failure to recognize/verify a signature was due to poor image quality and high similarity between two signatures. Recognition and verification ability of the system can be increased by using additional features in the input data set.