Table 2 Ablation study of Transformer Backbone, window shifted operation(WSO), loss function, downsampling structure, and hypernetwork structure.
From: Acquire continuous and precise score for fundus image quality assessment: FTHNet and FQS dataset
Type | Method | SRCC | PLCC | RMSE | Params(M) | FLOPS |
|---|---|---|---|---|---|---|
Backbones | ResNet41 | 0.1672 | 0.1630 | 39.24 | 27.97 | 12.70 G |
ConvNeXt-L42 | 0.8095 | 0.8081 | 12.19 | 55.25 | 26.30 G | |
MSG Transformer43 | 0.9241 | 0.9277 | 7.615 | 33.29 | 16.02 G | |
BTB | 0.9358 | 0.9442 | 7.024 | 14.88 | 6.044 G | |
WSO | w/o | 0.9263 | 0.9405 | 7.356 | 14.88 | 6.044 G |
with | 0.9358 | 0.9442 | 7.024 | 14.88 | 6.044 G | |
Loss | \(\mathscr {L}_1\) | 0.93245 | 0.9439 | 6.716 | 14.88 | 6.044 G |
\(\mathscr {L}_2\) | 0.9363 | 0.9389 | 7.028 | 14.88 | 6.044 G | |
\(\mathscr {L}_1 + \mathscr {L}_2\) | 0.93405 | 0.9408 | 6.914 | 14.88 | 6.044 G | |
\(\mathscr {L}_{smoothL1}\) | 0.9345 | 0.9447 | 6.581 | 14.88 | 6.044 G | |
Downsampling Structure | Direct | 0.9357 | 0.9426 | 6.978 | 0.572 | 82.39 M |
Stepwise | 0.9354 | 0.9437 | 6.988 | 0.393 | 56.67 M | |
Hypernetwork | w/o | 0.6092 | 0.6077 | 20.93 | 12.28 | 5.06 G |
with | 0.9358 | 0.9442 | 7.024 | 14.88 | 6.044 G |
