基于SRGAN的圖像超分辨率重建案例描述01任務案例目標02任務案例分析03任務案例實施04任務1案例描述1案例描述本案例基于SRGAN，并在T91數據集上訓練，實現圖像的超分辨重建。2案例目標案例目標學會搭建SRGAN超分辨率重建網絡；

通過約束對抗損失和內容損失訓練SRGAN 網絡以實現圖像的超分辨率重建。23案例分析案例分析SRGAN網絡同樣需要一些低分辨率圖像和原始高分辨率圖像，低分辨率圖像用做輸入數據來訓練SRGAN網絡模型，高分辨率圖片用來檢測超分圖像重建的效果。為此除了原始圖像，需要手動構造低分辨率圖像。整個案例的流程如下：（1）加載數據集，進行規(guī)范化；（2）構造低分辨率圖像；（3）搭建SRGAN網絡；（4）編譯并訓練；（5）預測并對比。34案例實施4案例實施 1.數據預處理創(chuàng)建utils.py，首先是一些預處理文件，包括讀取，存儲圖片等函數。#讀入圖片，并對圖片做尺寸裁減，并將所有圖片進行歸一化處理img=img[start_h:start_h+fine_size,start_w:start_w+fine_size,:]img=(img-127.5)/127.5#保存圖片以及圖片名稱defsave_images(images,size,filename):returnscipy.misc.imsave(filename,np.squeeze(images))4案例實施2.定義網絡層創(chuàng)建layer.py，主要定義一些特定的網絡模塊以及所需的損失函數。其中導入tensorflow.contrib.slim庫要求numpy版本為1.20.0以下，scipy版本為1.2.0。#定義殘差網絡塊defres_block(input_x,out_channels=64,k=3,s=1,scope='res_block'):withtf.variable_scope(scope):x=input_xinput_x=slim.conv2d_transpose(input_x,out_channels,k,s)input_x=slim.batch_norm(input_x,scope='bn1')input_x=tf.nn.relu(input_x)input_x=slim.conv2d_transpose(input_x,out_channels,k,s)input_x=slim.batch_norm(input_x,scope='bn2')returnx+input_x4案例實施#定義PSNR損失函數defPSNR(real,fake):mse=tf.reduce_mean(tf.square(127.5*(real-fake)+127.5),axis=(-3,-2,-1))psnr=tf.reduce_mean(10*(tf.log(255*255/tf.sqrt(mse))/np.log(10)))returnpsnr3.訓練模型

創(chuàng)建SRGAN.py，包含SRGAN模型的構建以及訓練測試過程。

定義生成器。生成器包含五個卷積模塊和五個殘差塊和兩個像素隨機操作。4案例實施conv1=tf.nn.relu(slim.conv2d_transpose(input_x,64,3,1,scope='g_conv1'))shortcut=conv1#res_block(input_x,out_channels=64,k=3,s=1,scope='res_block'):res1=res_block(conv1,64,3,1,scope='g_res1')res2=res_block(res1,64,3,1,scope='g_res2')res3=res_block(res2,64,3,1,scope='g_res3')res4=res_block(res3,64,3,1,scope='g_res4')res5=res_block(res4,64,3,1,scope='g_res5')conv2=slim.batch_norm(slim.conv2d_transpose(res5,64,3,1,scope='g_conv2'),scope='g_bn_conv2')conv2_out=shortcut+conv2#pixel_shuffle_layer(x,r,n_split):conv3=slim.conv2d_transpose(conv2_out,256,3,1,scope='g_conv3')shuffle1=tf.nn.relu(pixel_shuffle_layer(conv3,2,64))#64*2*2conv4=slim.conv2d_transpose(shuffle1,256,3,1,scope='g_conv4')shuffle2=tf.nn.relu(pixel_shuffle_layer(conv4,2,64))conv5=slim.conv2d_transpose(shuffle2,3,3,1,scope='g_conv5')self.g_vars=tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,'generator')4案例實施定義鑒別器。鑒別器包含四個卷積模塊和兩個全連接層。conv1=leaky_relu(slim.conv2d(input_x,64,3,1,scope='d_conv1'))conv1_1=leaky_relu(slim.batch_norm(slim.conv2d(conv1,64,3,2,scope='d_conv1_1'),scope='d_bn_conv1_1'))conv2=leaky_relu(slim.batch_norm(slim.conv2d(conv1_1,128,3,1,scope='d_conv2'),scope='d_bn_conv2'))conv2_1=leaky_relu(slim.batch_norm(slim.conv2d(conv2,128,3,2,scope='d_conv2_1'),scope='d_bn_conv2_1'))conv3=leaky_relu(slim.batch_norm(slim.conv2d(conv2_1,256,3,1,scope='d_conv3'),scope='d_bn_conv3'))conv3_1=leaky_relu(slim.batch_norm(slim.conv2d(conv3,256,3,2,scope='d_conv3_1'),scope='d_bn_conv3_1'))conv4=leaky_relu(slim.batch_norm(slim.conv2d(conv3_1,512,3,1,scope='d_conv4'),scope='d_bn_conv4'))conv4_1=leaky_relu(slim.batch_norm(slim.conv2d(conv4,512,3,2,scope='d_conv4_1'),scope='d_bn_conv4_1'))

conv_flat=tf.reshape(conv4_1,[self.batch_size,-1])dense1=leaky_relu(slim.fully_connected(conv_flat,1024,scope='d_dense1'))dense2=slim.fully_connected(dense1,1,scope='d_dense2')4案例實施定義優(yōu)化器，分為兩個優(yōu)化器，分別優(yōu)化生成器和鑒別器self.d_optim=tf.train.AdamOptimizer(learning_rate=self.learning_rate,beta1=self.beta1,beta2=self.beta2).minimize(self.d_loss,var_list=self.d_vars)self.g_optim=tf.train.AdamOptimizer(learning_rate=self.learning_rate,beta1=self.beta1,beta2=self.beta2).minimize(self.g_loss,var_list=self.g_vars)定義內容損失和對抗損失：content_loss=inference_mse_content_loss(real,fake)d