|
FER classification method | Algorithm | Advantage | Disadvantage |
|
Conventional machine learning algorithm | Viola-Jones algorithm | Fast face detection algorithm [34] | Has lower detection accuracy compared to more complex algorithms [34] |
Support vector machine (SVM) | High accuracy in classification tasks [77] | Computationally intensive, especially when dealing with large datasets or complex models [77] |
Support vector regression (SVR) | Less susceptible to overfitting [57] | Computationally expensive, particularly for large datasets [78] |
Extreme learning machine (ELM) | Perform faster in classification [59] | Takes more computational time and has a lower accuracy than PNN [59] |
Probabilistic neural network (PNN) | Takes less computational time than ELM and is more efficient in classification [59] | Possible overfitting of the data [59] |
Decision tree | Able to handle missing data by not incorporating the missing feature during the decision-making process [79] | More computation time compared to KNN [60] |
K-nearest neighbors (KNN) | Achieved higher accuracy when compared to the decision tree algorithm and lesser computation time [60] | Slower performance compared to decision tree [60] |
Multilayer perceptron (MLP) | Capable of adaptive learning and optimal processing [80] | Lower classification accuracy compared to the random forest algorithm [69] |
Adaptive boosting (AdaBoost) | Enhance the performance of classification out of weak learners [81] | Susceptible to noisy data [82] |
Random forest (RF) | Robust to noisy data or outliers [83] | Prone to overfitting [84] |
|
Deep learning algorithm | Convolutional neural networks (CNN) | Effective at handling complex image and video data [61] | Requires a large amount of training data and significant data augmentation to avoid overfitting [85] |
Neural network | Able to achieve high classification accuracy [33] | Long processing time [33] |
Deep belief network (DBN) | Robustness in classification [35] | Requires large amounts of training data [35] |
Long short-term memory (LSTM) | Addressed the issue of vanishing gradients [52] | Slow computational speed for complex architectures [52] |
Temporal relation network (TRN) | Able to achieve state-of-the-art performance on FER benchmarks [56] | Prone to overfitting/underfitting on small datasets [56] |
Deep facial spatiotemporal network (DFSTN) | Able to fuse facial spatiotemporal information [64] | Require larger amounts of training data to learn effective feature representations and to avoid overfitting [64] |
Deep CNN (DCNN) | Effective at learning complex features from raw image data [86] | Difficult to interpret, as it might be challenging to understand the underlying mechanisms behind the model’s decision-making process [87] |
Squeeze and excitation-deep adaptation networks (SE-DAN) | Can be used for transfer learning and domain adaptation [71] | Require a significant amount of computational resources and time to train [71] |
Multitask cascaded convolutional neural network (MTCNN) | High accuracy in detection and classification tasks; able to detect multiple faces in a single image [75] | Require a large amount of training data to achieve high accuracy [75] |
|
Hybrid algorithm | Support vector machine+convolutional neural network (SVM+CNN) | Enhance the performance of classification compared to using just one of these algorithms alone [30] | Hyperparameter tuning of this combination of two algorithms can be challenging and time-consuming [88] |
BERN (combination of temporal convolution, bidirectional LSTM, and attention mechanism) | Achieved state-of-art performance [47] | Requires a large amount of training data and a long training time [47] |
Hybrid convolutional neural network (hybrid CNN) | More robust to variations in input data [89] | Long training time [48] |
Hybrid deep neural network (hybrid DNN) | Able to handle a wide range of data types and classification tasks [90] | Requires a large amount of training data [54] |
|