Table 6 Strengths and gaps in traditional CNN models.
From: Improved facial emotion recognition model based on a novel deep convolutional structure
Accuracy | Strength | Gap | |
|---|---|---|---|
DenseNet121 | 74% | Reduce overfitting, increase feature reuse, reduce the number of parameters | The massive increase in parameters caused by an increase in the number of feature maps in every layer |
InceptionV3 | 77% | End up replacing large - sized filters with (1 × 7), and (1 × 5) filters | Complicated structure. Lack of consistency. |
VGG16 | 83% | Max Pooling has been applied | The use relatively demanding fully - connected layers |
VGG19 | 88% | More powerful than VGG16 | More complex, more computationally intensive potentially more prone to overfitting. |
ResNet50 | 87% | The concept of residual learning was introduced. Error rate for deeper networks has been lessened | Little-complex architecture |
Xception | 91% | More efficient in terms of processing time than standard convolutions. | Depth-wise separable convolution layer replaces the typical Inception modules |
EfficientNetB0 | 93% | Employs a compounded coefficient to scale all depth/width/resolution elements similarly | Slow training at the large image resolution |
