Table 7 Summary of ablation study impacts across all evaluated architectures, showing qualitative component contributions, vector size reduction percentage, and accuracy gains.
From: A multi-filter deep transfer learning framework for image-based autism spectrum disorder detection
Model | Image enhancement | Data augmentation | Scaling strategy | Vector reduction (%) | ACC gain (Test) | ACC gain (Val) |
|---|---|---|---|---|---|---|
AffectNet | Neutral (Test) | Positive (Test) | Standard/Robust (Val) | 79.3 | 3.33 | 5.00 |
AlexNet | Positive | Positive | Standard | 96.1 | 2.34 | 0.00 |
ResNet-50 | Positive | Negative | Standard | 97.0 | 1.33 | 8.00 |
VGG16 | Positive | Neutral/Negative | Neutral | 96.7 | 1.00 | 7.00 |
VGG19 | Positive | Neutral/Negative | Robust | 96.0 | 2.00 | 6.00 |
ViT | Positive | Neutral | MinMax | 72.2 | 3.00 | 4.00 |
ViTFER | Positive | Neutral | Standard | 73.5 | 1.33 | 6.00 |
ViTSwin | Strong Positive | Strong Positive | MinMax | 83.7 | 2.34 | 6.00 |
