summary by bard
ESRGAN is an enhanced super-resolution generative adversarial network (SRGAN) that achieves better perceptual quality than SRGAN. It proposes several improvements to the network architecture, discriminator, perceptual loss, and network interpolation.
Network architecture: ESRGAN removes all batch normalization (BN) layers from the generator and replaces the original residual block with the residual-in-residual dense block (RRDB). RRDB has a deeper and more complex structure than the original residual block, which helps to improve the perceptual quality of the reconstructed images.
Discriminator: ESRGAN replaces the standard discriminator with the relativistic average discriminator (RaD). RaD learns to predict the probability that a real image is more realistic than a fake image, which helps the generator to learn sharper edges and more detailed textures.
Perceptual loss: ESRGAN uses the features before activation rather than after activation as the perceptual loss. This helps to improve the brightness consistency and perceptual quality of the reconstructed images.
Network interpolation: ESRGAN proposes a network interpolation strategy to balance the perceptual quality and fidelity of the reconstructed images. This strategy can be used to continuously adjust the reconstruction style and smoothness.
ESRGAN was evaluated on the PIRM-SR Challenge. It achieved the best perceptual index in region 3, which is the region with the highest perceptual quality.
Overall, ESRGAN is an effective SRGAN that achieves better perceptual quality than SRGAN. It proposes several improvements to the network architecture, discriminator, perceptual loss, and network interpolation.
ESRGAN 是一种增强型超分辨率生成对抗网络 (SRGAN)，比 SRGAN 具有更好的感知质量。它提出了对网络架构、判别器、感知损失和网络插值的几个改进。
网络架构: ESRGAN 从生成器中移除所有批归一化 (BN) 层，并将原始残差块替换为残差-在-残差密集块 (RRDB)。RRDB 比原始残差块具有更深、更复杂的结构，有助于提高重建图像的感知质量。
判别器: ESRGAN 将标准判别器替换为相对平均判别器 (RaD)。RaD 学习预测真实图像比假图像更真实的概率，这有助于生成器学习更锐利的边缘和更详细的纹理。
感知损失: ESRGAN 使用激活之前而不是激活之后的特征作为感知损失。这有助于提高重建图像的亮度一致性和感知质量。
网络插值: ESRGAN 提出了一种网络插值策略来平衡重建图像的感知质量和保真度。该策略可用于连续调整重建风格和平滑度。
ESRGAN 在 PIRM-SR 挑战赛上进行了评估。它在区域 3 中获得了最佳感知指数，该区域是感知质量最高的区域。
总体而言，ESRGAN 是一种有效的 SRGAN，比 SRGAN 具有更好的感知质量。它提出了对网络架构、判别器、感知损失和网络插值的几个改进。
a far to do: riir
放大图片, 保留高频细节, 伪影
