Table 2.
Comparative analysis of countermeasure techniques
| Techniques | Purpose | Characteristics | Applicability | Advantages | Disadvantages |
|---|---|---|---|---|---|
| PCA [19, 27, 32, 34, 88, 149] | DR , FE | Unsupervised, maximize inter-class distance | PA, textural/structural PS, POcc | Noise reduction | Poor result for large dataset, outliers, higher variance |
| LDA [15, 19, 49, 65, 164, 186] | DR , FE | Supervised, maximize inter-class, minimize intra-class distance | PS, MU, PA, and ill | Reliable and efficient method | Assumption based method that can affect the results |
| SVD [62, 70, 188] | DR | Matrix decomposition, pattern-based solution, generalization of Eigen face | ill, CB | Optimized information using few coefficients | Slow, expensive singularity problem |
| DCV [4] | DR , CLS | Variation of Fisher’s LDA with small sample size | POcc, FS, Exp | Reduces singularity and small sample size problem | Handling of large matrices, complex method |
| Kernel PCA [170] | DR , FE | Transformation of non-linear patterns into linearly separable high-dimensional space | FR, handwriting recognition | Dealing with non-linear distribution-based unconstrained problem | Longer computation time, over fitting issue |
| kernel LDA [180] | DR | Non-parametric method, allows efficient computation of Fisher discriminant | FR, Exp | No assumption required for input distribution | Small sample size problem |
| DCT [6, 111, 117, 120, 174] | DR , CLS | Transformation-based holistic method used to represent the sum of sinusoids for different magnitudes and frequencies | FMo, textural and structural PS, ill | Fast, provides constant matrix, preserve energy | Quantization is required |
| LBP [29, 40, 46, 84, 96, 123, 176] | FE | Image texture-based analysis through spatial structure, mathematically proven | FSD, Ag, G | Robust, efficient for illumination, time, cost | Large false positive |
| HOGs [5, 31, 159, 186] [33] | FDe | Two main parameters, i.e., gradients direction and its magnitude | FSD, FMo, G, and illumination variations | Robust to variable lighting conditions | High dimensional feature space, cost, large datasets |
| SIFT [8, 9, 71, 142, 171] | FDe | Local features detection | PS, IT, POcc, LR, CB | Transformation invariance (S, Ro), efficient for Omni-directional | The complexity and run time |
| SURF [19, 60, 113, 144] | FDe | Extracts salient features (S, Orientation, ill) | POcc, LR | Eliminate the undesired motion found in videos, higher efficiency | Illumination variations issue |
| Gabor Wavelet [8, 26, 34, 56, 82] | FE, MM | Biological inspired features, scale and orientation based features | FSD, IT, Exp | Invariance to shift, rotate and illumination change | Large memory, cost, and higher dimensionality issue |
| Viola- Jones [13, 25, 71, 108, 129, 149] | FE , CLS | Robust and generalized technique for face recognition | LR, G | large features, fast, best for low-resolution images | Frontal face images required, sensitive to lighting conditions |
| Skin Color Modelling [76, 79, 146] | FE (Low-level) | Y parameter in YIQ, YUV and YCbCr shows the luminance, and other two for chrominance. Hue, saturation, and intensity contain the color depth, purity, and brightness, respectively | FD, MU, POcc | Depth color information, fast processing in controlled environment | Not suitable for unconstrained condition, performance is dependent on the color-model used |
| SVM [10, 13, 20, 25, 27, 51, 135, 138, 151, 184] | CLS | Multi-class classifier, support vectors, decision boundary, and kernel discriminative classifier | Structural PS, FSD, FMo, POcc, and Exp | Handle noise, less chance of over fitting, real valued features | High computational cost |
| K-NN algorithm [4, 15, 40, 176] | CLS | An alternate of SVM, unsupervised clustering-based | FMo, Exp | Suitable to find out the loss/error estimation | Not fit for large dataset, long process time |
| HMM [112, 132] | CLS | A generative classifier focused on sequence of symbol emitted by system underlying random walk between states | Pattern recognition, classification, and structure analysis | Strong statistical foundation | Not suitable for higher order correlation |
| SLNN [2] | FE , CLS | Human brain oriented feed forward neural network consist of two layer architecture | POcc | Easy setup and less computation | Separable data is desired, cannot deal complex non-linear problems efficiently |
| MLNN [119, 156] | Automatic FE , CLS | At least one hidden layer is required including input and output layer | FSD, Exp, medical diagnosis | Easily tackle complex problem | Heavy computation, large space, long time |
| CNNs [52, 123, 124, 154, 167, 177, 178, 183, 184] | Image-based FE , CLS | A deep learning technique, which takes image data as input | FSD, Po, Exp, Occ, LR | Supports transfer learning by sharing the pre-trained weights, fast | Layers interpretations is not clear, complicated hidden layer mechanism |
| Euclidean-DMC [9, 13, 45, 112] | CLS , LF | Distance between two data sample (lets p and q) for n-dimensional feature space | PS, G, LR, Occ | An effective method to find uniqueness | Assumes in prior for misplacing of data points |
| Manhattan- DMC [82] | CLS , LF | Distance between two data samples measured along the axes at right angle | FR, video surveillance, Crime monitoring, Occ | This method has robustness to outliers | Generates large value for two similar images that represents the dissimilarity |
| Chi-Square- DMC [176] | CLS | Investigates the difference between what actually find in study (observed), and what is expected to find (hypothesis) | Histogram matching, Exp | Suitable for comparing different histograms, easy computation and interpretation | Requires data is in numeric form to deal with higher degree of categories |
| Cosine Similarity-DMC [45, 111, 117] | CLS, SE | Similarity index between two different vectors, Cosine angular represents the product of two vectors with direction | PS, FR and camera orientation | Good accuracy, vectors are used to measure similarity, direction and angular displacement | If two vectors lies on the same line than the cosine value will be 1, and the similarity value will be 0 |
DR-Dimesionality Reduction, FE- Feature Extraction, FDe- Feature Descriptor, FR- Face Recognition, FD- Face Detection, FS- Face Spoofing Detection, CLS-Classification, PA- Presentation Attack, PS- Plastic Surgery, MU- Makeup, MM-Multiple Modularity, LF- Loss Function, IT- Identical Twins, FMo- Face Morphing, POcc-Partial Occlusion, Exp- Expression, Ag-Aging, R- Race, G- Gender, Po- Pose, ill- Illumination variations, LR- Low Resolution, CB- Cluttered Background, CO-Camera Orientation, S-Scaling, Ro-Rotation, DMC-Distance Metric Classifier, SE- Similarity Evaluation