Research Article
A Convolutional Self-Attention Network for CSI Reconstruction in MIMO System
Table 5
NMSE (dB), cosine similarity
, and FLOPs, where CR is the compression ratio.
| | CR | Method | Indoor | Outdoor | FLOPs | | NMSE | | NMSE | |
| | 1/4 | CsiNet | -17.36 | 0.99 | -8.75 | 0.91 | 5.41 M | | LightCNN | -18.41 | 0.99 | -9.31 | 0.92 | 24.72 M | | CRNet | -24.10 | / | -12.57 | / | 24.57 M | | CLNet | -29.16 | / | -12.88 | / | 4.05 M | | CSANet-const | -30.66 | 0.99 | -13.80 | 0.94 | 45.83 M | | CSANet-cosine | -34.18 | 0.99 | -14.73 | 0.95 | 45.83 M |
| | 1/16 | CsiNet | -8.65 | 0.93 | -4.51 | 0.79 | 3.84 M | | LightCNN | -12.06 | 0.95 | -6.11 | 0.86 | 21.58 M | | CRNet | -10.52 | / | -5.36 | / | 23.00 M | | CLNet | -11.15 | / | -5.73 | / | 2.48 M | | CSANet-const | -13.79 | 0.94 | -6.29 | 0.86 | 44.26 M | | CSANet-cosine | -15.54 | 0.96 | -6.52 | 0.82 | 44.26 M |
| | 1/32 | CsiNet | -6.24 | 0.89 | -2.81 | 0.67 | 3.58 M | | LightCNN | -9.92 | 0.93 | -3.05 | 0.69 | 21.35 M | | CRNet | -8.90 | / | -3.16 | / | 22.74 M | | CLNet | -8.95 | / | -3.49 | / | 2.22 M | | CSANet-const | -10.21 | 0.93 | -4.14 | 0.78 | 43.99 M | | CSANet-cosine | -11.24 | 0.94 | -4.45 | 0.79 | 43.99 M |
| | 1/64 | CsiNet | -5.84 | 0.87 | -1.93 | 0.59 | 3.45 M | | LightCNN | -3.97 | 0.79 | -2.27 | 0.65 | 21.02 M | | CRNet | -6.23 | / | -2.19 | / | 22.61 M | | CLNet | -6.34 | / | -2.19 | / | 2.09 M | | CSANet-const | -6.14 | 0.86 | -2.41 | 0.65 | 43.86 M | | CSANet-cosine | -6.56 | 0.88 | -2.86 | 0.67 | 43.86 M |
|
|
