SRGAN超分辨重建

最新推荐文章于 2024-05-20 00:52:19 发布
最新推荐文章于 2024-05-20 00:52:19 发布
阅读量1.2k 收藏 7
点赞数
分类专栏： 深度学习
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/bblingbbling/article/details/107295010
版权

SRGAN超分辨重建

简介

SRGAN全称：super-resolution generative adversarial network，是利用生成对抗网络进行图像超分辨率重建的一种深度学习网络。

效果展示

在这里插入图片描述

网络结构

网络由生成网络和判别网络组成。

生成网络：input + Conv(channel = 64, stride = 1) + Relu + B个残差块 + 2次Deconv(4倍放大)
判别网络：input + Conv(channel = 64, stride = 1) + Leaky Relu + 重复卷积块(Conv+Leak Relu+BN)进行下采样 + Dense + Sigmoid (结果和1或0对比)
在这里插入图片描述

Loss函数

1.优化目标
在这里插入图片描述
2.损失函数
总损失 = 内容损失 + 对抗损失 + 正则项
其中，正则项基于全变分损失，用来去噪，使图像变得平滑。
在这里插入图片描述
3.感知损失
感知损失 = 内容损失 + 对抗损失

内容损失：
采用VGG提取特征，计算fake和gt特征的欧氏距离。
在这里插入图片描述
尽管使用传统的MSE损失函数能实现高的信噪比PSNR，但缺失了高频内容，使生成的图像过于平滑。所以使用VGG损失函数。

对抗损失：
判别生成网络生成的样本。
在这里插入图片描述

训练过程

生成网络和判别网络交替训练。

判别网络：

# original的imgs_hr和valid=1对比
# generator的fake_hr和fake=0对比

imgs_hr, imgs_lr = self.data_loader.load_data(batch_size)
fake_hr = self.generator.predict(imgs_lr)

d_loss_real = self.discriminator.train_on_batch(imgs_hr, valid=1)
d_loss_fake = self.discriminator.train_on_batch(fake_hr, fake=0)
d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)

生成网络：

# generator的fake_hr判别结果validity和valid=1对比
# 通过VGG的fake_features和image_features对比

fake_hr = self.generator(img_lr)
fake_features = self.vgg(fake_hr)
validity = self.discriminator(fake_hr)
self.combined = Model(img_lr, [validity, fake_features])

imgs_hr, imgs_lr = self.data_loader.load_data(batch_size)

image_features = self.vgg.predict(imgs_hr)
g_loss = self.combined.train_on_batch(imgs_lr, [valid=1, image_features])

网络计算参数

判别网络参数

input = generate (512, 512, 3)--------> output = (32, 32, 1) 
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_3 (InputLayer)         (None, 512, 512, 3)       0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 512, 512, 64)      1792      
_________________________________________________________________
leaky_re_lu_1 (LeakyReLU)    (None, 512, 512, 64)      0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 256, 256, 64)      36928     
_________________________________________________________________
leaky_re_lu_2 (LeakyReLU)    (None, 256, 256, 64)      0         
_________________________________________________________________
batch_normalization_1 (Batch (None, 256, 256, 64)      256       
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 256, 256, 128)     73856     
_________________________________________________________________
leaky_re_lu_3 (LeakyReLU)    (None, 256, 256, 128)     0         
_________________________________________________________________
batch_normalization_2 (Batch (None, 256, 256, 128)     512       
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 128, 128, 128)     147584    
_________________________________________________________________
leaky_re_lu_4 (LeakyReLU)    (None, 128, 128, 128)     0         
_________________________________________________________________
batch_normalization_3 (Batch (None, 128, 128, 128)     512       
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 128, 128, 256)     295168    
_________________________________________________________________
leaky_re_lu_5 (LeakyReLU)    (None, 128, 128, 256)     0         
_________________________________________________________________
batch_normalization_4 (Batch (None, 128, 128, 256)     1024      
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 64, 64, 256)       590080    
_________________________________________________________________
leaky_re_lu_6 (LeakyReLU)    (None, 64, 64, 256)       0         
_________________________________________________________________
batch_normalization_5 (Batch (None, 64, 64, 256)       1024      
_________________________________________________________________
conv2d_7 (Conv2D)            (None, 64, 64, 512)       1180160   
_________________________________________________________________
leaky_re_lu_7 (LeakyReLU)    (None, 64, 64, 512)       0         
_________________________________________________________________
batch_normalization_6 (Batch (None, 64, 64, 512)       2048      
_________________________________________________________________
conv2d_8 (Conv2D)            (None, 32, 32, 512)       2359808   
_________________________________________________________________
leaky_re_lu_8 (LeakyReLU)    (None, 32, 32, 512)       0         
_________________________________________________________________
batch_normalization_7 (Batch (None, 32, 32, 512)       2048      
_________________________________________________________________
dense_1 (Dense)              (None, 32, 32, 1024)      525312    
_________________________________________________________________
leaky_re_lu_9 (LeakyReLU)    (None, 32, 32, 1024)      0         
_________________________________________________________________
dense_2 (Dense)              (None, 32, 32, 1)         1025      
=================================================================
Total params: 5,219,137
Trainable params: 5,215,425
Non-trainable params: 3,712

生成网络参数

input = original (128, 128, 3)--------> output = (512, 512, 3) 
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_4 (InputLayer)            (None, 128, 128, 3)  0                                            
__________________________________________________________________________________________________
conv2d_9 (Conv2D)               (None, 128, 128, 64) 15616       input_4[0][0]                    
__________________________________________________________________________________________________
activation_1 (Activation)       (None, 128, 128, 64) 0           conv2d_9[0][0]                   
__________________________________________________________________________________________________
conv2d_10 (Conv2D)              (None, 128, 128, 64) 36928       activation_1[0][0]               
__________________________________________________________________________________________________
batch_normalization_8 (BatchNor (None, 128, 128, 64) 256         conv2d_10[0][0]                  
__________________________________________________________________________________________________
activation_2 (Activation)       (None, 128, 128, 64) 0           batch_normalization_8[0][0]      
__________________________________________________________________________________________________
conv2d_11 (Conv2D)              (None, 128, 128, 64) 36928       activation_2[0][0]               
__________________________________________________________________________________________________
batch_normalization_9 (BatchNor (None, 128, 128, 64) 256         conv2d_11[0][0]                  
__________________________________________________________________________________________________
add_1 (Add)                     (None, 128, 128, 64) 0           batch_normalization_9[0][0]      
                                                                 activation_1[0][0]               
__________________________________________________________________________________________________
conv2d_12 (Conv2D)              (None, 128, 128, 64) 36928       add_1[0][0]                      
__________________________________________________________________________________________________
batch_normalization_10 (BatchNo (None, 128, 128, 64) 256         conv2d_12[0][0]                  
__________________________________________________________________________________________________
activation_3 (Activation)       (None, 128, 128, 64) 0           batch_normalization_10[0][0]     
__________________________________________________________________________________________________
conv2d_13 (Conv2D)              (None, 128, 128, 64) 36928       activation_3[0][0]               
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 128, 128, 64) 256         conv2d_13[0][0]                  
__________________________________________________________________________________________________
add_2 (Add)                     (None, 128, 128, 64) 0           batch_normalization_11[0][0]     
                                                                 add_1[0][0]                      
__________________________________________________________________________________________________
conv2d_14 (Conv2D)              (None, 128, 128, 64) 36928       add_2[0][0]                      
__________________________________________________________________________________________________
batch_normalization_12 (BatchNo (None, 128, 128, 64) 256         conv2d_14[0][0]                  
__________________________________________________________________________________________________
activation_4 (Activation)       (None, 128, 128, 64) 0           batch_normalization_12[0][0]     
__________________________________________________________________________________________________
conv2d_15 (Conv2D)              (None, 128, 128, 64) 36928       activation_4[0][0]               
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 128, 128, 64) 256         conv2d_15[0][0]                  
__________________________________________________________________________________________________
add_3 (Add)                     (None, 128, 128, 64) 0           batch_normalization_13[0][0]     
                                                                 add_2[0][0]                      
__________________________________________________________________________________________________
conv2d_16 (Conv2D)              (None, 128, 128, 64) 36928       add_3[0][0]                      
__________________________________________________________________________________________________
batch_normalization_14 (BatchNo (None, 128, 128, 64) 256         conv2d_16[0][0]                  
__________________________________________________________________________________________________
activation_5 (Activation)       (None, 128, 128, 64) 0           batch_normalization_14[0][0]     
__________________________________________________________________________________________________
conv2d_17 (Conv2D)              (None, 128, 128, 64) 36928       activation_5[0][0]               
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 128, 128, 64) 256         conv2d_17[0][0]                  
__________________________________________________________________________________________________
add_4 (Add)                     (None, 128, 128, 64) 0           batch_normalization_15[0][0]     
                                                                 add_3[0][0]                      
__________________________________________________________________________________________________
conv2d_18 (Conv2D)              (None, 128, 128, 64) 36928       add_4[0][0]                      
__________________________________________________________________________________________________
batch_normalization_16 (BatchNo (None, 128, 128, 64) 256         conv2d_18[0][0]                  
__________________________________________________________________________________________________
activation_6 (Activation)       (None, 128, 128, 64) 0           batch_normalization_16[0][0]     
__________________________________________________________________________________________________
conv2d_19 (Conv2D)              (None, 128, 128, 64) 36928       activation_6[0][0]               
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 128, 128, 64) 256         conv2d_19[0][0]                  
__________________________________________________________________________________________________
add_5 (Add)                     (None, 128, 128, 64) 0           batch_normalization_17[0][0]     
                                                                 add_4[0][0]                      
__________________________________________________________________________________________________
conv2d_20 (Conv2D)              (None, 128, 128, 64) 36928       add_5[0][0]                      
__________________________________________________________________________________________________
batch_normalization_18 (BatchNo (None, 128, 128, 64) 256         conv2d_20[0][0]                  
__________________________________________________________________________________________________
activation_7 (Activation)       (None, 128, 128, 64) 0           batch_normalization_18[0][0]     
__________________________________________________________________________________________________
conv2d_21 (Conv2D)              (None, 128, 128, 64) 36928       activation_7[0][0]               
__________________________________________________________________________________________________
batch_normalization_19 (BatchNo (None, 128, 128, 64) 256         conv2d_21[0][0]                  
__________________________________________________________________________________________________
add_6 (Add)                     (None, 128, 128, 64) 0           batch_normalization_19[0][0]     
                                                                 add_5[0][0]                      
__________________________________________________________________________________________________
conv2d_22 (Conv2D)              (None, 128, 128, 64) 36928       add_6[0][0]                      
__________________________________________________________________________________________________
batch_normalization_20 (BatchNo (None, 128, 128, 64) 256         conv2d_22[0][0]                  
__________________________________________________________________________________________________
activation_8 (Activation)       (None, 128, 128, 64) 0           batch_normalization_20[0][0]     
__________________________________________________________________________________________________
conv2d_23 (Conv2D)              (None, 128, 128, 64) 36928       activation_8[0][0]               
__________________________________________________________________________________________________
batch_normalization_21 (BatchNo (None, 128, 128, 64) 256         conv2d_23[0][0]                  
__________________________________________________________________________________________________
add_7 (Add)                     (None, 128, 128, 64) 0           batch_normalization_21[0][0]     
                                                                 add_6[0][0]                      
__________________________________________________________________________________________________
conv2d_24 (Conv2D)              (None, 128, 128, 64) 36928       add_7[0][0]                      
__________________________________________________________________________________________________
batch_normalization_22 (BatchNo (None, 128, 128, 64) 256         conv2d_24[0][0]                  
__________________________________________________________________________________________________
activation_9 (Activation)       (None, 128, 128, 64) 0           batch_normalization_22[0][0]     
__________________________________________________________________________________________________
conv2d_25 (Conv2D)              (None, 128, 128, 64) 36928       activation_9[0][0]               
__________________________________________________________________________________________________
batch_normalization_23 (BatchNo (None, 128, 128, 64) 256         conv2d_25[0][0]                  
__________________________________________________________________________________________________
add_8 (Add)                     (None, 128, 128, 64) 0           batch_normalization_23[0][0]     
                                                                 add_7[0][0]                      
__________________________________________________________________________________________________
conv2d_26 (Conv2D)              (None, 128, 128, 64) 36928       add_8[0][0]                      
__________________________________________________________________________________________________
batch_normalization_24 (BatchNo (None, 128, 128, 64) 256         conv2d_26[0][0]                  
__________________________________________________________________________________________________
activation_10 (Activation)      (None, 128, 128, 64) 0           batch_normalization_24[0][0]     
__________________________________________________________________________________________________
conv2d_27 (Conv2D)              (None, 128, 128, 64) 36928       activation_10[0][0]              
__________________________________________________________________________________________________
batch_normalization_25 (BatchNo (None, 128, 128, 64) 256         conv2d_27[0][0]                  
__________________________________________________________________________________________________
add_9 (Add)                     (None, 128, 128, 64) 0           batch_normalization_25[0][0]     
                                                                 add_8[0][0]                      
__________________________________________________________________________________________________
conv2d_28 (Conv2D)              (None, 128, 128, 64) 36928       add_9[0][0]                      
__________________________________________________________________________________________________
batch_normalization_26 (BatchNo (None, 128, 128, 64) 256         conv2d_28[0][0]                  
__________________________________________________________________________________________________
activation_11 (Activation)      (None, 128, 128, 64) 0           batch_normalization_26[0][0]     
__________________________________________________________________________________________________
conv2d_29 (Conv2D)              (None, 128, 128, 64) 36928       activation_11[0][0]              
__________________________________________________________________________________________________
batch_normalization_27 (BatchNo (None, 128, 128, 64) 256         conv2d_29[0][0]                  
__________________________________________________________________________________________________
add_10 (Add)                    (None, 128, 128, 64) 0           batch_normalization_27[0][0]     
                                                                 add_9[0][0]                      
__________________________________________________________________________________________________
conv2d_30 (Conv2D)              (None, 128, 128, 64) 36928       add_10[0][0]                     
__________________________________________________________________________________________________
batch_normalization_28 (BatchNo (None, 128, 128, 64) 256         conv2d_30[0][0]                  
__________________________________________________________________________________________________
activation_12 (Activation)      (None, 128, 128, 64) 0           batch_normalization_28[0][0]     
__________________________________________________________________________________________________
conv2d_31 (Conv2D)              (None, 128, 128, 64) 36928       activation_12[0][0]              
__________________________________________________________________________________________________
batch_normalization_29 (BatchNo (None, 128, 128, 64) 256         conv2d_31[0][0]                  
__________________________________________________________________________________________________
add_11 (Add)                    (None, 128, 128, 64) 0           batch_normalization_29[0][0]     
                                                                 add_10[0][0]                     
__________________________________________________________________________________________________
conv2d_32 (Conv2D)              (None, 128, 128, 64) 36928       add_11[0][0]                     
__________________________________________________________________________________________________
batch_normalization_30 (BatchNo (None, 128, 128, 64) 256         conv2d_32[0][0]                  
__________________________________________________________________________________________________
activation_13 (Activation)      (None, 128, 128, 64) 0           batch_normalization_30[0][0]     
__________________________________________________________________________________________________
conv2d_33 (Conv2D)              (None, 128, 128, 64) 36928       activation_13[0][0]              
__________________________________________________________________________________________________
batch_normalization_31 (BatchNo (None, 128, 128, 64) 256         conv2d_33[0][0]                  
__________________________________________________________________________________________________
add_12 (Add)                    (None, 128, 128, 64) 0           batch_normalization_31[0][0]     
                                                                 add_11[0][0]                     
__________________________________________________________________________________________________
conv2d_34 (Conv2D)              (None, 128, 128, 64) 36928       add_12[0][0]                     
__________________________________________________________________________________________________
batch_normalization_32 (BatchNo (None, 128, 128, 64) 256         conv2d_34[0][0]                  
__________________________________________________________________________________________________
activation_14 (Activation)      (None, 128, 128, 64) 0           batch_normalization_32[0][0]     
__________________________________________________________________________________________________
conv2d_35 (Conv2D)              (None, 128, 128, 64) 36928       activation_14[0][0]              
__________________________________________________________________________________________________
batch_normalization_33 (BatchNo (None, 128, 128, 64) 256         conv2d_35[0][0]                  
__________________________________________________________________________________________________
add_13 (Add)                    (None, 128, 128, 64) 0           batch_normalization_33[0][0]     
                                                                 add_12[0][0]                     
__________________________________________________________________________________________________
conv2d_36 (Conv2D)              (None, 128, 128, 64) 36928       add_13[0][0]                     
__________________________________________________________________________________________________
batch_normalization_34 (BatchNo (None, 128, 128, 64) 256         conv2d_36[0][0]                  
__________________________________________________________________________________________________
activation_15 (Activation)      (None, 128, 128, 64) 0           batch_normalization_34[0][0]     
__________________________________________________________________________________________________
conv2d_37 (Conv2D)              (None, 128, 128, 64) 36928       activation_15[0][0]              
__________________________________________________________________________________________________
batch_normalization_35 (BatchNo (None, 128, 128, 64) 256         conv2d_37[0][0]                  
__________________________________________________________________________________________________
add_14 (Add)                    (None, 128, 128, 64) 0           batch_normalization_35[0][0]     
                                                                 add_13[0][0]                     
__________________________________________________________________________________________________
conv2d_38 (Conv2D)              (None, 128, 128, 64) 36928       add_14[0][0]                     
__________________________________________________________________________________________________
batch_normalization_36 (BatchNo (None, 128, 128, 64) 256         conv2d_38[0][0]                  
__________________________________________________________________________________________________
activation_16 (Activation)      (None, 128, 128, 64) 0           batch_normalization_36[0][0]     
__________________________________________________________________________________________________
conv2d_39 (Conv2D)              (None, 128, 128, 64) 36928       activation_16[0][0]              
__________________________________________________________________________________________________
batch_normalization_37 (BatchNo (None, 128, 128, 64) 256         conv2d_39[0][0]                  
__________________________________________________________________________________________________
add_15 (Add)                    (None, 128, 128, 64) 0           batch_normalization_37[0][0]     
                                                                 add_14[0][0]                     
__________________________________________________________________________________________________
conv2d_40 (Conv2D)              (None, 128, 128, 64) 36928       add_15[0][0]                     
__________________________________________________________________________________________________
batch_normalization_38 (BatchNo (None, 128, 128, 64) 256         conv2d_40[0][0]                  
__________________________________________________________________________________________________
activation_17 (Activation)      (None, 128, 128, 64) 0           batch_normalization_38[0][0]     
__________________________________________________________________________________________________
conv2d_41 (Conv2D)              (None, 128, 128, 64) 36928       activation_17[0][0]              
__________________________________________________________________________________________________
batch_normalization_39 (BatchNo (None, 128, 128, 64) 256         conv2d_41[0][0]                  
__________________________________________________________________________________________________
add_16 (Add)                    (None, 128, 128, 64) 0           batch_normalization_39[0][0]     
                                                                 add_15[0][0]                     
__________________________________________________________________________________________________
conv2d_42 (Conv2D)              (None, 128, 128, 64) 36928       add_16[0][0]                     
__________________________________________________________________________________________________
batch_normalization_40 (BatchNo (None, 128, 128, 64) 256         conv2d_42[0][0]                  
__________________________________________________________________________________________________
add_17 (Add)                    (None, 128, 128, 64) 0           batch_normalization_40[0][0]     
                                                                 activation_1[0][0]               
__________________________________________________________________________________________________
up_sampling2d_1 (UpSampling2D)  (None, 256, 256, 64) 0           add_17[0][0]                     
__________________________________________________________________________________________________
conv2d_43 (Conv2D)              (None, 256, 256, 256 147712      up_sampling2d_1[0][0]            
__________________________________________________________________________________________________
activation_18 (Activation)      (None, 256, 256, 256 0           conv2d_43[0][0]                  
__________________________________________________________________________________________________
up_sampling2d_2 (UpSampling2D)  (None, 512, 512, 256 0           activation_18[0][0]              
__________________________________________________________________________________________________
conv2d_44 (Conv2D)              (None, 512, 512, 256 590080      up_sampling2d_2[0][0]            
__________________________________________________________________________________________________
activation_19 (Activation)      (None, 512, 512, 256 0           conv2d_44[0][0]                  
__________________________________________________________________________________________________
conv2d_45 (Conv2D)              (None, 512, 512, 3)  62211       activation_19[0][0]              
==================================================================================================
Total params: 2,042,691
Trainable params: 2,038,467
Non-trainable params: 4,224
bblingbbling
关注
  • 0
    点赞
  • 踩
  • 7
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
专栏目录
SRGAN图像超分辨率重建
10-18 336
SRGAN的目标是将低分辨率（LR）图像转换为高分辨率（HR）图像，以提高图像的视觉质量和细节。生成器通过欺骗判别器来生成更逼真的高分辨率图像，而判别器则不断学习区分真实图像和生成图像，以提供更准确的反馈。SRGAN在图像超分辨率重建领域取得了显著的成果，能够生成更清晰、更真实的高分辨率图像，对于图像增强、电视和视频内容的提升等应用具有重要意义。判别器则负责区分生成器生成的图像和真实高分辨率图像。生成器负责将输入的低分辨率图像映射到高分辨率图像空间，并生成类似于真实高分辨率图像的输出。
基于飞桨复现 SRGAN 模型，对图像进行超分辨率重构
百度大脑
10-28 1699
点击左上方蓝字关注我们【飞桨开发者说】侯继旭，海南师范大学自动化本科在读，PPDE飞桨开发者技术专家，研究方向为目标检测、对抗生成网络等SRGAN是一种用于图像超分辨率（SR）的生成对抗...
图像超分辨率重建相关概念、评价指标、数据集、模型
最新发布
a486259的博客
05-20 2275
超分辨率（Super-Resolution），简称超分（SR）。是指利用光学及其相关光学知识，根据已知图像信息恢复图像细节和其他数据信息的过程，简单来说就是增大图像的分辨率，使图像更加“清晰”，尽可能保证图像质量不下降。超分辨率的方法包括传统方法和深度学习的方法，有关超分辨率方法分类如下图所示。深度学习方法在性能上远远领先于传统方法，有着更好的图像超分辨率表现。知乎用户 阿布的足迹​ https://zhuanlan.zhihu.com/p/558813267从单一低分辨率图像中恢复高分辨率图像。
【Super Resolution】超分辨率——SRGAN
PRIS-SCMonkey的博客
08-27 6565
接触这篇paper的理由——据说这是第一篇将GAN应用到超分领域的论文。在SRGAN之前，个人认为，超分网络的本质 Paper：Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network ...
SRGAN超分辨率网络
不要给自己设限，尝试更多可能（思所向 皆可往）
04-17 1888
超分辨率论文阅读—SRGAN - 知乎提示：以下是本篇文章正文内容，下面案例可供参考1、超分辨率问题由低清图像恢复的高清图像采用的是MSE(Mean Square Error)作为损失函数，该损失函数会造成恢复出来的图像高频信息不足，视觉感知不佳。使用MSE训练的网络恢复出的图像高频信息不足，整体图像趋于模糊。基于此问题，SRGAN提出使用GAN将恢复的图像拟合到真实数据集的分布上。论文中有一幅图很好的解释了区别：超分辨率是一个病态问题（ill-posed)，一个低清图像块可以对应多个高清图像块。
SRGAN——使用与超分辨率重建的GAN
新手的博客
05-20 6034
  SRGAN数据GAN理论在超分辨率重建（SR）方面的应用。 一、超分辨率技术 1.SR技术介绍   SR技术，是指从观测到的低分辨率图像重建出相对应的高分辨率图像，在监控设备、卫星图像和医学影像等领域都有重要的应用价值，也可以应用于马赛克图片的恢复应用场景。   SR可分为两类：从多张低分辨率图像重建出高分辨率图像；从单张低分辨率图像重建出高分辨率图像。基于深度学习的SR，主要是基于单张低分辨率的重建方法，即Single Image Super-Resolution(SISR)。   SISR是一个逆问
论文笔记——ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks
TracelessLe的专栏
01-28 5372
基本信息 标题：ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks 简称：ESRGAN 时间：2018年初版，1 Sep 2018, ECCV2018 PIRM Workshop 作者：Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Chen Change Loy, Yu Qiao, Xiaoou Tang； CUHK-SenseTime Jo
SRGAN超分辨率图像重建
03-26
SRGAN是超分辨率图像重建领域的创新性工作，它通过引入生成对抗网络，不仅提升了图像的分辨率，还增强了图像的视觉质量和细节表现。随着深度学习技术的不断发展，我们可以期待更多高效、高质量的SR方法出现，进一步...
SRGAN_SRGAN_生成对抗网络_超分辨率重建_超分辨率
09-11
在“SRGAN_SRGAN_生成对抗网络_超分辨率重建_超分辨率”这个主题中，我们将深入探讨一种特别有效的SR方法——生成对抗网络（Generative Adversarial Networks，简称GANs）以及其在超分辨率领域的应用——SRGAN。...
超分辨率重建SRGAN代码（包括将SRGAN生成器单独拿出来做残差网络）
03-05
超分辨率重建技术是一种重要的图像处理方法，主要用于提升低分辨率图像的质量，使其接近或达到高分辨率图像的清晰度。SRGAN（Super-Resolution Generative Adversarial Networks）是该领域的一个里程碑式工作，由...
基于GAN的手写体数字超分辨率重建 掌握SRGAN重建方法，对MNIST图像超分辨率重建
08-20
《基于GAN的手写体数字超分辨率重建：理解SRGAN在MNIST图像上的应用》 超分辨率重建技术在图像处理领域中扮演着至关重要的角色，它能够将低分辨率（LR）图像提升至高分辨率（HR）状态，从而提高图像的细节和清晰度...
超分辨重建毕业设计，主要基于ESRGAN和SRGAN实现超分重建模型
11-08
<项目介绍> 该资源内项目源码是个人的毕设，代码都测试ok，都是运行成功后才上传资源，答辩评审平均分达到96分，放心下载使用！ 1、该资源内项目代码都经过测试运行成功，功能ok的情况下才上传的，请放心下载使用！...
ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks, ECCV2018
heitang1802的博客
07-05 695
ESRGAN任务是超分，是ECCV18年的文章，曾在PIRM2018超分竞赛获得第一名，模型是在SRGAN基础上进行的改进；而SRGAN(CVPR17)是最早用GAN来做超分的文章，其2个特色分别是： （1）生成器使用残差块。使用残差结构主要是由于设计的网络比较深，为了保证梯度信息的有效传播，增强网络稳定性。 （2）相较同行较早使用perceptual loss，是不是它最早提出的有待考证。 回到ESRGAN，它在SRGAN基础上把模型的几个组成成分都改了：（1）生成器结构改进。（2）判别器改用.
ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks 翻译
黄笳倞的博客
01-13 1万+
ESRGAN：增强型超分辨率生成对抗网络 项目地址：https://github.com/xinntao/BasicSR 摘要 超分辨率生成对抗网络（SR GAN）[1]是一项开创性的工作，能够在单图像超分辨率期间生成逼真的纹理。 然而，幻觉细节通常伴随着令人不快的伪影。 为了进一步提高视觉质量，我们深入研究了SRGA...
对抗生成网络学习（七）——SRGAN生成超分辨率影像(tensorflow实现)
QQ704630835的博客
11-13 1万+
一、背景 SRGAN(Super-Resolution Generative Adversarial Network)即超分辨率GAN，是Christian Ledig等人于16年9月提出的一种对抗神经网络。利用卷积神经网络实现单影像的超分辨率，其瓶颈仍在于如何恢复图像的细微纹理信息。对于GAN而言，将一组随机噪声输入到生成器中，生成的图像质量往往较差。因此，作者提出了SRGAN，并定义一个lo...
【图像超分辨率重建】——SRGAN/SRResNet论文精读笔记
Sunny_Coder的博客
11-10 3782
SR的不是适定性导致SR图像缺乏纹理细节，有监督的SR算法的优化目标通常是最小化恢复的HR图像和地面真实之间的平均平方误差（MSE）。这很方便，因为最小化MSE也能使峰值信噪比（PSNR）最大化，而峰值信噪比是用来评估和比较SR算法的一个常用指标。然而，MSE（和PSNR）捕捉感知上相关差异的能力是非常有限的，比如高纹理细节，因为它们是基于像素级的图像差异来定义的。最高的PSNR并不一定反映出感知上更好的SR结果。
SRGAN简单了解
读书郎
09-01 1万+
超分辨率问题的病态性质尤其表现在取较高的放大因子时，重构的超分辨率图像通常会缺失纹理细节。监督SR算法的优化目标函数通常取重建高分辨率图像和地面真值之间的均方误差，在减小均方误差的同时又可以增大峰值信噪比(PSNR)，PSNR是评价和比较SR算法的常用指标。但是MSE和PSNR值的高低并不能很好的表示视觉效果的好坏。正如Figture2表现出的，PSNR最高并不能反映最好的视觉SR效果。 SRGAN，使用了结合跳跃-连接（skip-connection）的深度残差网络（ResNet）。我们使用VGG网络的高
SRGAN的损失函数
weixin_38203533的博客
06-21 9379
GAN的作用，是额外增加一个D网络和2个损失（判别损失和生成判别损失），用一种交替训练的方式训练两个网络。这个模型可以分为3部分：main模块，adversarial模块，和vgg模块。（一般main模块就是adversarial模块里的G网络）adversarial可以看作是一种训练技巧，只在训练阶段会用到adversarial模块进行计算，而在推断阶段，仅仅使用G网络（或者说main模块）。也...
图像的超分辨率重建SRGAN与ESRGAN
热门推荐
qxqsunshine的博客
12-23 1万+
SRGAN 传统的图像超分辨率重建方法一般都是放大较小的倍数，当放大倍数在4倍以上时就会出现过度平滑的现象，使得图像出现一些非真实感。SRGAN借助于GAN的网络架构生成图像中的细节。 训练网络使用均方误差（MSE）能够获得较高的峰值信噪比（PSNR），但是恢复出来的图像会丢失图像的高频细节信息，在视觉上有不好的体验感。SRGAN利用感知损失（perceptual loss）和对抗损失（adv...
SRGAN实现超分辨率重建
06-09
SRGAN是一种基于生成对抗网络的超分辨率模型，其核心思路是通过训练一个生成器网络和一个判别器网络，使得生成器网络能够将低分辨率图像转换为高分辨率图像，并且判别器网络能够区分生成的高分辨率图像和真实高分辨率图像之间的差异。 具体实现上，SRGAN模型包含两个部分：生成器网络和判别器网络。生成器网络由一系列卷积层、反卷积层和残差块组成，用于将低分辨率图像转换为高分辨率图像。判别器网络则由一系列卷积层和全连接层组成，用于判别生成的高分辨率图像和真实高分辨率图像之间的差异。 在训练过程中，生成器网络和判别器网络相互博弈，不断优化模型参数，直到最终生成的高分辨率图像能够和真实高分辨率图像无法区分为止。具体来说，训练过程包括以下几个步骤： 1. 准备训练数据。将高分辨率图像和对应的低分辨率图像作为训练数据，其中低分辨率图像可以通过对高分辨率图像进行下采样得到。 2. 训练判别器网络。首先，使用真实高分辨率图像和生成器网络生成的高分辨率图像作为正样本，使用下采样的高分辨率图像作为负样本，训练判别器网络，使其能够区分正样本和负样本。 3. 训练生成器网络。首先，使用生成器网络将低分辨率图像转换为高分辨率图像，然后使用判别器网络判别生成的高分辨率图像的质量。通过最小化生成的高分辨率图像和真实高分辨率图像之间的差异，以及最大化判别器网络判别生成的高分辨率图像为真实高分辨率图像的概率，来优化生成器网络的参数。 4. 微调生成器网络。在训练过程中，生成器网络可能会出现过拟合或者欠拟合的情况，需要通过微调生成器网络的参数来解决这些问题。 SRGAN模型的训练需要大量的高分辨率图像和低分辨率图像作为训练数据，并且需要在GPU上进行训练，因此需要一定的计算资源和时间。但是，SRGAN模型能够生成非常逼真的高分辨率图像，对于一些对图像质量要求较高的应用场景具有很大的帮助。

“相关推荐”对你有帮助么？

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值