Lowest memory consumption and second shortest runtime in NTIRE 2022 on Efficient Super-Resolution.
- Enhanced Residual Block.
- High-Frequency Attention Block.
- Batch Normalization layers can be applied to attention branch to boost performance.
Our goal is to design a strightforward but powerful backbone for lightweight image super-resolution, so the testing model topology is really simple (only contains five highly optimized operators: 3x3 convolution, LeakyReLU, element-wise addition, element-wise multiplication and sigmoid).
Since there are no other tricks, you can directly adopt EDSR framework to train the model.