Abstract


1. Introduction


2. Method

2.1 The architecture of FilterNet

Image

2.2. Training data

2.3 Real-time inference process

Image


3. Performance analysis of FilterNet

3.1 Single data analysis

3.2. The accuracy of FilterNet

$$ \mathrm{PSNR} = 20 \cdot \log_{10} \left( \frac{I_{\max}}{\mathrm{RMSE}(x, d)} \right) $$

$x$ 表示模型输出波形;
$d$ 表示标签波形;
$\mathrm{RMSE}(x, d)$ 为均方根误差;
$I_{\max}$ 为最大值(通常归一化为 1);

$$ \mathrm{SSIM}(x, d) = \frac{(2 \mu_x \mu_d + C_1)(2 \sigma_{xd} + C_2)}{(\mu_x^2 + \mu_d^2 + C_1)(\sigma_x^2 + \sigma_d^2 + C_2)} $$

$\mu_x, \mu_d$ 为均值;
$\sigma_x^2, \sigma_d^2$ 为方差;
$\sigma_{xd}$ 为协方差;
$C_1, C_2$ 为常数项(用于稳定计算);

3.3. Computation Efficiency Test

3.4. Model Generalization Ability Test


4. Disscussion

Image


5. Conclusion

转载请注明出处