|
| Author | Dataset | Features extraction techniques | Classification techniques | Problems | Results |
|
| Hamadène et al. [41] off-line | CEDAR dataset | Contourlet transform (CT) and cooccurrence matrix features | Support vector machines (SVM) classifier | Feature extraction methods do not allow capturing contours contained into an image. | (1) AER of 0.07 for writer dependent approach |
| (2) AER of 0.18 for writer independent approach |
|
| Nemmour and Chibani [6] off-line | CEDAR dataset | Ridgelet transform and grid features | Support vector machines (SVM) classifier | The system can achieve higher accuracies but requires larger runtime. | EER is equal to 4.18 |
|
| Kamihira et al. [42] Combined | Collected signature from 19 persons, 798 genuine samples and 684 skilled forgeries samples | Gradient features | Support vector machine (SVM) classifier | Few signature samples will increase the FRR of genuine signatures while too many samples will be labor intensive for the user. | Accuracy is equal to 97.22% |
|
| Griechisch et al. [43] online | SigComp2011 | x, y coordinates, pressure, and velocity features | Kolmogorov-Smirnov distribution distance | Some reference signature which differed the most from the other reference were excluded, so it was not used during the decision process | EER is less than 13%. |
|
| Fayyaz et al. [20] online | SVC2004 | Method based on learned signature features using autoencoder classifier | One-class classifiers | The system has been designed base on one hidden layer. | EER is equal to 2.15 |
|
| Radhika and Gopika [4] combined | The dataset used is collected from 13 different writers. For each person 30 genuine and 25 forged signatures are collected. | (1) Pen tip tracking features were utilized in online case | Support vector machines (SVM) classifier | | (1) FAR is equal to 11.54 |
| (2) Gradient and projection features were utilized in off-line case | (2) FRR is equal to 34.62 |
| | (3) AER is equal to 23.08 |
|
| Lech and Czyzewski [44] online | Collected signatures using a wacom tablet from 10 persons for each person 5 signatures | Static features and time-domain functions of signals | Dynamic time warping (DTW) | The main drawback associated with using a graphical tablet with no display is lack of the visual feedback while putting down a signature. | |
|
| Hamadene and Chibani [45] off-line | (1) CEDAR | Contourlet transform (CT) based directional code Co-occurrence matrix (DCCM) technique. | Writer-independent decision thresholding | The verification step is performed using only the feature dissimilarity measure | (1) AER for CEDAR is 2.10 |
| (2) GPDS | (2) AER for GPDS is 18.42 |
| Taşkiran and Çam [46] off-line | Collected signature images at Yildiz technical university from 15 person, 40 sample from each. | Histogram of oriented gradients (HOG) features | Generalized regression neural networks (GRNN) algorithm | Large implementation costs and processing time | Accuracy is equal to 98.33% |
|
| Suryani et al. [47] off-line | They use 80 samples of signatures obtained from 8 persons | Moment invariant features | Efficient fuzzy Kohonen clustering network (FKCN) algorithm. | The accuracy of the training data is smaller than the accuracy of the test data. | Accuracy is equal to 70% |
|
| Serdouk et al. [37] off-line | (1) CEDAR | Histogram of template (HOT) features | Support vector machine (SVM) classifier | Highlight strokes orientation in handwritten signatures. | (1) For CEDAR AER is 1.03%. |
| (2) MCYT-75 | (2) For MCYT-75 AER is 6.40% |
|
| Sharif et al. [38] off-line | (1) CEDAR | Global and local features selected using genetic algorithm | Support vector machine (SVM) classifier | High error rate | (1) AER for CEDAR is 4.67 |
| (2) MCYT | (2) AER for MCYT is 5.0 |
| (3) GPDS | (3) AER for GPDS is 3.75 |
|
| Antal et al. [36] online | (1) MOBISIG | (1) Function-based system use local features | (1) In function-based system DTW was utilized for distance calculation among the test signature and the reference signatures. | Feature-based methods offer poor results in the case of global threshold. | (1) EER is equal to 0.01% for random forgeries and 5.81% for skilled forgeries when user-specific thresholds is used. |
| (2) DOODB | (2) Feature based system use global features | (2) In feature-based system euclidean distance used in training and manhattan distance is used in testing. | (2) EER is equal to 1.68% fort random forgeries and 14.31% for skilled forgeries when global thresholds is used. |
|
| Jia et al. [21] online | SVC2004 Task2 | (1) Shape context features | SC-DTW was used to compare the test signature with the all the reference signatures | Require more training samples and consumes more computation costs. | EER is equal to 2.39% |
| (2) Function–based features |
|
Mersa et al. [48]
off-line | (1) MCYT | Convolutional neural network (CNN) | Support vector machine (SVM) classifier | Deep networks need rich and plentiful training data, which is rare in signature datasets. | (1) UTsig EER is 9.80% |
| (2) UTsig | (2) MYCT EER is 3.98% |
| (3) GPDS-synthetic | (3) GPDS-synthetic EER is 6.81% |
|
| Saleem and Kovari [19] online | (1) MCYT-100 | (1) Horizontal position | DTW and sampling frequency for each signer | External factors may affect the accuracy of the results | Accuracy improved in about 70% of the total 500 tests and 92% in the chosen system. |
| (2) SVC2004 | (2) Vertical position |
| (3) SigComp’11 (Dutch) | (3) Pressure |
| (4) SigComp’11 (Chinese) | (4) Horizontal and vertical positions combination |
| (5) SigComp’13 (Japanese) | (5) Horizontal position, vertical position, and pressure combination |
|
| Semih et al. [49] online | The dataset was created from scratch and examples were collected from 40 persons. | Time sequential peak values were used to construct the feature vector. | Dynamic time warping (DTW) algorithm | Classification difficulty caused by different paper types pen types and phone models. | (1) EER vary between %8.14–%16.61 when signer-specific thresholds is used. |
| (2) EER vary between %15.29–%28.45 when single threshold for all signers is used. |
|
| Foroozandeh et al. [50] off-line | (1) GPDS-synthetic | (1) Circlet transform (CT) | (1) Support vector machine (SVM) | The proposed method did not outperform on MYCT-75 dataset. | (1) EER with GDS-synthetic is 5.67 |
| (2) MYCT-75 | (2) Statistical properties was calculated by the gray level co-occurrence matrices (GLCM) | (2) k-Nearest neighbor (k-NN) | (2) EER with MYCT-75 is 7 when r = 1 and 8.20 when r = 10 |
| (3) UTSig | (3) EER with YTSig is 6.72 |
|
| Bonde et al. [51] off-line | (1) GPDS | Fine-tuned CNN was used as signature features extraction technique | Support vector machine (SVM) classifier | | (1) Accuracy for GPDS is 92.03 |
| (2) MYCT-75 | (2) Accuracy for MYCT-75 is 90.78 |
| (3) UTSig | (3) Accuracy for UTSig is 85.46 |
|
| Kurowski et al. [52] online | Hand-corrected dataset containing 10,622 signatures were obtained with electronic pen | Convolutional neural network to extract meaningful features from signatures | Deep convolutional network, the triplet loss method was used to train a neural network | The algorithm used in the proposed method becomes progressively slower as the training procedure continues. | (1) EER is equal to 5.77% for random forgery |
| (2) 11.114% EER for skilled forgery |
|
| Zhou et al. [17] combined | Collected off-line images and online data of 1200 signatures | (1) For offline features the texture and geometric features are extracted using GLCM and HOG | (1) Support vector machine (SVM) classifier | Large intra-class and inter-class variability. | (1) Accuracy for SVM is 81.17% |
| (2) Dynamic time warping (DTW) | (2) Accuracy for DTW is 89.17% |
| (2) For online features velocity, acceleration, angle and radius of curvature are extracted | (3) SF-A | (3) Accuracy for SF-A is 93.08% |
| (4) SFL | (4) Accuracy for SF-L is 92.58% |
|
| Melhaoui and Benchaou [53] off-line | Collected dataset from 12 person, 20 signature from each | Histogram of oriented gradients (HOG) features | Fuzzy min max classification (FMMC) method | The recognition rate depends highly on the choice of the sensitivity parameter which regulates how fast the membership value decreases. | Recognition rate is equal to 96% |
|
