Signature Recognition based on Discrete Wavelet Transform

1.1. Signatures Recognition Systems

Handwritten signatures are simply divided into two categories:

1.2. Levels of Forgeries

Forgeries generally, based on making, forging, or fixing nearly any document, with the intent to different kinds of forgeries are classified into the subsequent categories [14]:

1.2.3. Skilled forgery

Undoubtedly, the foremost tough of all forgeries is formed by skilled cheaters or persons who have expertise in signature repetition. To detect this, one may either trace or imitate the signature by arduous manner. Various types of forgeries and related original signatures are shown in Fig. 1 [17].

Fig. 1. Types of forgery.

1.3. Signature Characteristics

The signature should be treated as an image, because a person can use any symbol, letter, group of letters, etc., like signature. Therefore, it is possible not to separate the written letters, or maybe you cannot understand what is written. Therefore, the applied system for the analysis of signatures must use image processing ideas. Most likely, the signature of a signer may vary for each sign based on many factors. However, there must be a unique feature to identify the signature so that it can be used as biometrics. Below some of the essential features [18]:

As signature conjointly belongs to at least one of the parameters that satisfy the characteristics; therefore, it is going to be used as a signal of person identification.

1.4. Literature Review

There are many methods that used for signature recognition; this section explains some of these methods:

Bharadi and Kekre (2010) showed a comprehensive review of handwriting recognition systems and special attention is paid to the analysis of static signature recognition systems (SRSs). They discussed a set of conventional features based on offline SRS, as well as cluster-based global characteristics. This system is multi-algorithmic; such systems combine the benefits of individual feature sets and improve recognition rates. The system reported a precision of 95.08%, which is higher than the individual performance measures [19].

Tomar and Singh (2011) proposed a method to compare the performance of energy density methods; the directional characteristic method and the proposed directional method with the energy method, based on the time required for training, accuracy, acceptance, and rejection ratios. The obtained results showed that if the samples available for training are limited, then the method of directional characteristics is better than the density of the energy method but the merging of the directional characteristics with the energy method, leading to the best method. While the time required for training is a bit longer for this proposed method, accuracy and false acceptance rate (FAR) are satisfactory, and this additional time may be bearable. The obtained results indicated that False Rejection Ratio (FRR) varies randomly for all cases. In addition, this research gives the reason for future studies [20].

Radmehr et al. (2012) demonstrated that signature recognition poses a problem of recognition of difficult model. This research introduced a new SRS based on the transformation of Radon, shape dimension (fractal dimension), and support vector machine (SVM). Signature verification competition (SVC) database was used to evaluate the performance of this method. The results of the simulation indicate that the planned technique is promising in recognition of signatures. In addition, the performance of the SVM classification has been evaluated with different cores. This research presented an effective algorithm, and it may be valid in a large signature database where many signature styles are taken into account. Other attractive perspectives and challenges for future research are based on two aspects: How to extract more efficient features and how to combine an SVM-based classifier with other signature recognition methods [21].

Kanawade and Katariya (2013) explained that a person’s signature is an important biometric attribute of a human being that can be used to authenticate human identity. However, human signatures can be treated as an image and recognized by artificial vision. They examined the problem of personal authentication through the recognition of signatures. Online and offline methods have been described in this work. The method of signature verification indicated that pre-processing and future extraction limits the advantage of being extremely acceptable to potential customers over the rest of the biometric solutions. This research indicated that verification and recognition of the offline identity are provided each time the signature is captured and delivered to the user in an image format. The signatures that are verified can support the parameters extracted from the signature using many image processing techniques. This research presented a case study of the signature verification method for offline signature verification [22].

Hiary et al. (2013) implemented an Arabic and Persian offline signature verification system. This technique is divided mainly into four stages: Pre-processing, recording of images, extraction of features and finally, verification of signatures. Pre-processing operations were performed to reduce processing time and increase accuracy. Then applied discrete wavelet transform (DWT) to extract the characteristics of the signature image and used logical operations with mathematical formulas to verify the signatures. The experimental results have shown reasonable improvements in many recent works. This approach reduces the number of levels of DWT levels and the amount of training required, with a low percentage of FAR of 1.56% and a false rejection rate of 10.9% and a low average error rate of 6.23%. Future plans include improving this technique to increase the accuracy of verification and testing techniques with different forms and types of signatures [23].

Kanawade and Katariya (2013) explained a method to verify the handwritten signature in offline mode. This approach proposed a new technique of recognition and verification of signatures in offline mode based on global characteristics, mask, signature grid, and SVM. It is better to try counterfeiting using the online signature verification method. The concept of offline signature verification can be extended to other biometric identification systems, such as handwriting. Combined signature features with other biometric aspects such as speech and facial recognition; it provides better results than any individual system [24].

Bhattacharya et al. (2013) implemented a signature authentication machine to provide a security system of simple, safe, and fast biometric behavior. Based on some mathematical coordinates, this methodology is faster than the different forms. Color technique is applied to make it safer. The interface of this application is extremely simple, which makes it easy to use and easy user-friendly. The implemented system has some limitations: It can only identify static changes in a signature and cannot identify any dynamic change in a signature. This signature authentication system works only for offline mode, and the results were tested only offline. The proposed system for recognizing and verifying offline signatures is implemented through pixel matching technique (PMT). PMT is used to verify the signature of the user with the signature stored in the database. The performance of the proposed method was compared with the existing back propagation artificial neural network and SVM [25].

Doroz et al. (2014) proposed an efficient method of verification of handwritten signatures. This method creates complex functionalities for each signature that describe this signature. These characteristics are based on the analysis of the dependencies between the dynamic characteristics recorded by the tablets. These complex features are used to create vectors that describe the signature. The elements of these vectors are calculated using the measures indicated in this research. The similarity between the signatures was evaluated by determining the similarity of the vectors in the comparative signatures. The obtained results showed a high effectiveness of the verification using the proposed method where the precision obtained was 96.67% [26].

Malik and Arora (2015) talked about two types of signature verification system are present offline (static) and online (dynamic). The biometrics has an important advantage over traditional authentication techniques due to the fact that biometric characteristics of the individual are not easily acceptable for convert, are unique of every person, and cannot be lost, stolen, or broken. The signature is acquired as an image; this image shows a personal style of human handwriting. A signature verification system sometimes focuses on the detection of one or more type of forged signatures. The purpose of this work is to design and develop an offline handwritten signature verification system that will differentiate between the original signature and closer forge signature depending on some special features [27].

Al-Ani and Al-Saidi (2011) studied how to improve handwritten Arabic signature recognition. They this work proposed SRS used for many applications and have played an important role in these applications. This work presented a simple and efficient signature recognition approach based on the discrete Fourier transform used to extract features. Many people are registered through a built-in signature database. The results obtained indicated that a good and effective recognition rate was recognized, as the results showed, it provides value almost greater than zero [28].

Hafemann et al. (2015) proposed a method for offline signature verification. Although the categorization of authentic signatures and qualified counterfeits remains a difficult task, the error rates have decreased significantly in recent years, mainly due to the progress of deep learning applied to the task used for the verification of Persian firms of convolutional neural network (CNN). Analyzing the recent contributions to the field, they can notice that they concentrate in the following categories: Obtaining better features, improving classification with a limited number of samples, augmenting the datasets, and building model ensembles. As a suggestion for future work, this research can be continued to explore higher feature representations and introducing methods to improve classification with a limited number of samples [29].

Hedjaz et al. (2018) proposed an offline signature recognition method using binary features. It is clear; the offline signatures recognition is mainly important for the authentication of administrative and official documents that required greater precision. The characteristics are extracted using two descriptors: Binary statistical image and local binary models. To evaluate the reliability of the method, experiments were conducted using MCYT-75 and GPDS-100 databases. Using a nearest k-neighbor classifier, the recognition performance reaches a high 97.3% and 96.1% for the MCYT-75 and GPDS-100 databases, respectively. During the verification of the signature, the accuracy of the classification measured with the same error index equal error rate (EER) reached, respectively, 4.2% and 4.8% in the GPDS-100 and the GPDS-160. In addition, the EER for the MCYT-75 database reached 7.78% [30].

1.5. Statement of the Problem

Many researches are published regarding signature recognition issue but many of these researched suffered from many aspects. This paper presents an offline SRS. DWT is suggested to extract common features and different change of the original image and signal of signature. We expected an efficient result in signature recognition that supports the verification step.

2.1. Data Collection

The offline signature images are acquired from traditional papers. This requires a high specific signature image that can be achieved through high-resolution scanner. In general, the first step of an offline signature verification system is to extract these signatures from traditional papers using a scanner. In this research, signatures are collected using either black or blue ink on a white A4 paper. Each A4 paper is divided into 10 parts to cover 10 signatures per page. A group of 100 persons are used to collect 10 signatures from each person, which resulting 1000 signature samples as shown in Fig. 2.

Fig. 2. Sample of signatures.

2.2. Implemented Approach

The offline signature approach for verification and recognition is divided into many steps as shown in Fig. 3:

Fig. 3. Signature verification and recognition system.

3.1. Pre-processing

Pre-processing step plays an important role in signature recognition to overcome any raised problem. The purpose of this step is to make standard signatures that are ready for the feature extraction process. In this research, the pre-processing step is achieved through the following operations.

3.1.1. Noise removal

Minimizing of noise is an important operation that affects the overall signature recognition process because the signature image may contain unwanted information. In this research, the implemented noise removal operation is performing through filtering that not affects the details of signature image information. The median filter is an adequate filter that resolves the problems of noise reduction. Both signature images before and after applying media filter are shown in Figs. 4 and 5, respectively.

Fig. 4. Signature image before noise removal.

Fig. 5. Signature image after noise removal.

3.1.2. Image filtering and binarization

The true color image red green blue (RGB) is converted into grayscale intensity image through eliminating the hue and saturation information while retaining the luminance element. A grayscale image could be an information matrix whose values represent intensities within some range wherever every part of the matrix corresponds to at least one image pixel. The process of signature image filtering and binarization is shown in Figs. 6 and 7, respectively.

Fig. 6. Signature image before binarization.

Fig. 7. Signature image after binarization.

3.1.3. Width normalization

Signature dimensions could have intrapersonal and interpersonal variations. That the image width is adjusted to a default value and also the heights can be modified considering constant height-to-width ratio, as a result, the width dimension is adjusted to 100. Signature image before and after normalization is shown in Figs. 8 and 9, respectively.

Fig. 8. Signature image before normalization.

Fig. 9. Signature image after normalization.

3.1.4. Thinning

The goal of the thinning operation is to eliminate the thickness variations of a pen by creating the line of one pixel of thickness that leading to accurate recognition rate. Furthermore, signatures are thinned so as to reduce the computations needed by the graph matching algorithm. Figs. 10 and 11 represent signature images before and after thinning.

Fig. 10. Signature image before thinning.

Fig. 11. Signature image after thinning.

3.2. Registration

The registration step contains two main operations; scaling and shifting. In scaling, the signature is re-scaled into the suitable type to realize the higher and accurate result. After that, shifting operation is applied to determine the center of the signature. These operations are mentioned below [5].

3.2.1. Scaling

This step resizing all the signature images for all persons. Alternative techniques performed this resizing supported the maximum height and width for all samples in the database. This leading to unnecessary, superfluous stretching of signatures and may cause errors in verification. To avoid this problem, the implemented technique uses the maximum height and width for signatures of the same person rather than unifying over all dataset where Figs. 12 and 13 show the signature images before and after scaling.

Fig. 12. Signature image before scaling.

Fig. 13. Signature image after scaling.

3.2.2. Shifting

In this step, the scaled signature image is aligned to the center of gravity of each person’s signatures. Figs. 14 and 15 show the signature images before and after applying to the align process.

Fig. 14. Signature image before alignment.

Fig. 15. Signature image after alignment.

3.3. Feature Extraction

Feature extraction is an important step to enhance the signature image characteristics that are necessary to apply the comparison operation. DWT is to extract the perfect features before the verification step. The advantage of using vectors is the fact that they always have the same length, which does not depend on the size of the signature.

DWT produces multi-scale image decomposition using filtering and sub-sampling that leading to the shape of the decomposition image where terribly effectively revealing knowledge redundancy in many scales.

Since signature images are two-dimensional (2D) signals, so it is better to increase the theme to 2D space by applying the transformed row and column-wise, respectively. According to 2D DWT, four sub bands are generated for each level of DWT. One low-pass sub-band containing the coarse approximation of the source image referred to as low-low (LL) sub-band, and three high-pass sub-bands that exploit image details across totally different directions, high-low for horizontally, low-high for vertical, and high-high (HH) for diagonal details, as shown in Fig. 16.

Fig. 16. Fig. 16. Discrete wavelet transform bands: Low-low, low-high, high-low, and high-high.

Applying DWT, each sample of the signature image is decomposed into four images, the primary image represents the LL band values, whereas the other three images represent the HH in vertical, diagonal, and horizontal directions, respectively, as shown in Fig. 17. DWT is especially used to extract the features from the signature image. The proposed technique uses high pass images to extract the mandatory data for signature verification.

Fig. 17. Discrete wavelet transform, (A1) approximation results, (V1) vertical, (H1) horizontal, and (D1) diagonal high pass results.

The DWT of a signal is calculated by passing it through a series of filters. First the samples are passed through a low-pass filter with impulse response resulting in a convolution of the two:

The signal is additionally rotten at the same time using a high-pass filter. The outputs are given the brief coefficients (from the high-pass filter) and approximation coefficients (from the low-pass). It is necessary that the two filters are associated with one another and that they are called a construction mirror filter.

Each sample is changed during the decomposition first level and decomposition second level, the primary image represents the low pass values, whereas the opposite three images represent the first level high pass in vertical, diagonal, and horizontal directions, and the second level of primary image represents the low pass values whereas the opposite three images represent the second level high pass in vertical, diagonal, and horizontal directions. Fig. 18 illustrates the differences between the two levels of the same sample.

Fig. 18. sDifferent between two levels of discrete wavelet transform.

Fig. 19 represents two levels reconstructed signature image compared with the original image.

Fig. 19. Original image and 2-level reconstruct the image.

Fig. 20 shows the obtained first-level one side DWT of a noisy Doppler signal using the wavelet (high pass) and compared that with the scaling (low pass) filter.

Fig. 20. First-level one-side signal filtering and scaling.

Fig. 21 shows the reconstructed of the smoothed version of the signal using the approximation coefficients compared with the original signal, in which a high similarity is obtained.

Fig. 21. Original signal compared with reconstruct signal.