RDN

Metric.log

@@ -0,0 +1,75 @@
+/scratch/gilbreth/li3120/RDB/Test_Result/Manga109_16.h5
+average ssim:0.957323
+average psnr:32.212730
+/scratch/gilbreth/li3120/RDB/Test_Result/Set14_19.h5
+average ssim:0.865827
+average psnr:28.640444
+/scratch/gilbreth/li3120/RDB/Test_Result/Set14_17.h5
+average ssim:0.945727
+average psnr:34.091388
+/scratch/gilbreth/li3120/RDB/Test_Result/Manga109_17.h5
+average ssim:0.983521
+average psnr:36.844997
+/scratch/gilbreth/li3120/RDB/Test_Result/Set14_18.h5
+average ssim:0.867282
+average psnr:28.860134
+/scratch/gilbreth/li3120/RDB/Test_Result/B100_18.h5
+average ssim:0.815428
+average psnr:26.904180
+/scratch/gilbreth/li3120/RDB/Test_Result/Urban100_19.h5
+average ssim:0.820174
+average psnr:25.668797
+/scratch/gilbreth/li3120/RDB/Test_Result/Urban100_20.h5
+average ssim:0.785363
+average psnr:24.468557
+/scratch/gilbreth/li3120/RDB/Test_Result/Urban100_18.h5
+average ssim:0.828697
+average psnr:25.981574
+/scratch/gilbreth/li3120/RDB/Test_Result/Set5_16.h5
+average ssim:0.961617
+average psnr:34.722022
+/scratch/gilbreth/li3120/RDB/Test_Result/Set5_18.h5
+average ssim:0.940866
+average psnr:32.595460
+/scratch/gilbreth/li3120/RDB/Test_Result/B100_20.h5
+average ssim:0.757086
+average psnr:25.223548
+/scratch/gilbreth/li3120/RDB/Test_Result/Manga109_19.h5
+average ssim:0.921437
+average psnr:29.053814
+/scratch/gilbreth/li3120/RDB/Test_Result/Manga109_20.h5
+average ssim:0.880391
+average psnr:26.505380
+/scratch/gilbreth/li3120/RDB/Test_Result/B100_16.h5
+average ssim:0.866474
+average psnr:28.625940
+/scratch/gilbreth/li3120/RDB/Test_Result/B100_19.h5
+average ssim:0.826464
+average psnr:27.226383
+/scratch/gilbreth/li3120/RDB/Test_Result/Set5_17.h5
+average ssim:0.980510
+average psnr:38.318403
+/scratch/gilbreth/li3120/RDB/Test_Result/Manga109_18.h5
+average ssim:0.925166
+average psnr:29.275654
+/scratch/gilbreth/li3120/RDB/Test_Result/Urban100_16.h5
+average ssim:0.884011
+average psnr:28.174515
+/scratch/gilbreth/li3120/RDB/Test_Result/Set5_19.h5
+average ssim:0.938823
+average psnr:32.228928
+/scratch/gilbreth/li3120/RDB/Test_Result/Urban100_17.h5
+average ssim:0.942854
+average psnr:32.116574
+/scratch/gilbreth/li3120/RDB/Test_Result/Set14_16.h5
+average ssim:0.901079
+average psnr:30.622271
+/scratch/gilbreth/li3120/RDB/Test_Result/Set14_20.h5
+average ssim:0.811752
+average psnr:26.307208
+/scratch/gilbreth/li3120/RDB/Test_Result/Set5_20.h5
+average ssim:0.885235
+average psnr:28.487348
+/scratch/gilbreth/li3120/RDB/Test_Result/B100_17.h5
+average ssim:0.927497
+average psnr:31.904156

Metric.py

@@ -0,0 +1,31 @@
+import numpy as np
+import h5py
+import os
+from glob import glob
+import matplotlib.pyplot as plt
+from skimage.transform import rescale
+from skimage.measure import compare_psnr,compare_ssim
+
+paths=glob('/scratch/gilbreth/li3120/RDB/Test_Result/*')
+
+for path in paths:
+    print(path)
+    f= h5py.File(path, "r")
+    rec=f['rec']
+    gt=f['gt']
+    Ave_ssim=np.zeros((len(gt),1))
+    Ave_psnr=np.zeros((len(gt),1))
+    def rgb2ycbcr (img):
+        y = 16 + (65.481 * img[:, :, 0]) + (128.553 * img[:, :, 1]) + (24.966 * img[:, :, 2])
+        return y / 255
+    for i in range(len(gt)):
+        img=np.squeeze(rec[i,:,:,:])
+        imgg=np.squeeze(gt[i,:,:,:])
+        img=np.clip(img,a_min=0,a_max=1)
+        img=rgb2ycbcr(img)
+        imgg=rgb2ycbcr(imgg)
+        Ave_ssim[i]=compare_ssim(img,imgg)
+        Ave_psnr[i]=compare_psnr(img,imgg)
+    print('average ssim:%f'%(np.mean(Ave_ssim)))
+    print('average psnr:%f'%(np.mean(Ave_psnr)))
+    f.close()

RDB_run.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+module use /depot/itap/amaji/modules
+module spider learning
+module load learning/conda-5.1.0-py36-gpu
+module load ml-toolkit-gpu/tensorflow
+module load ml-toolkit-gpu/keras
+module load  ml-toolkit-gpu/opencv
+module list
+source activate /home/li3120/.conda/envs/cent7/5.1.0-py36/tfgan1
+cd /home/li3120/ECE570
+python /home/li3120/ECE570/model.py

README.md

@@ -1 +1,43 @@
-# RDN-SR-Keras
+# course-project-DamonLee5
+# Acknowledgement
+For the best of my knowledge, my project is the first RDN implementation written by Keras. However, I still need to acknowledge that https://github.com/yulunzhang/RDN and https://github.com/hengchuan/RDN-TensorFlow help me for some hyperparameter setting.
+
+# Data preparation
+## utils.py
+This file generates our training datasets for three different degradations, including 5 datasets.
+  - Bicubic degradation with factor 2, 3, and 4.
+  - BD degradation with factor 3
+  - DN degradation with factor 3
+Also this file generates 25 datasets for 5 degradations on 5 benchmark datasets, which have not been seen while training.
+To generate those dataset, I wrote function to chop, rescale, filter, read, and save the image.
+
+# DNN Model
+I implement this project by Keras. I utilize a standard Keras framework with a data_loader and model file.
+## data_loader.py
+This file is used for load the data from the dataset for training, sampling and testing.
+## model.py
+This file include the entire RDN network. Also, it can test on validation set while traing and load a trained model for testing.
+
+# Training
+RDB_run.sh is used to train in cluster by qsub.
+
+# Test Result Generation
+## test.py
+Since we need to calculate the metric, and online calculate takes too long, I first generate the entire reconstructed high resolution dataset for metric calculation. This result dataset include ground truth and reconstructed image.
+
+## Metric.py
+Calculate the Metric among all 25 testing datasets. Result is in the Metric.log, which is consistent with the result displayed in the term paper.
+
+## loss_plot.ipynb
+Plot loss from training log file.
+
+## Utils_test.ipynb
+Generate an entire high resolution image from reconstructed patches.
+
+# Saved model
+https://drive.google.com/open?id=1sQZO1ZZ4MIYM_vF5wOm6e9F8OiyeoF_H
+- RDB_best_16.h5, BI degradation, factor 3
+- RDB_best_17.h5, BI degradation, factor 2
+- RDB_best_18.h5, BI degradation, factor 4
+- RDB_best_19.h5, BD degradation, factor 3
+- RDB_best_20.h5, DN degradation, factor 3

data_loader.py

@@ -0,0 +1,92 @@
+import scipy
+from glob import glob
+import numpy as np
+import matplotlib.pyplot as plt
+import cv2
+from skimage.io import imread
+from skimage import data_dir
+from skimage.transform import radon, rescale
+from skimage.transform import iradon
+import h5py
+class DataLoader():
+    def __init__(self, dataset_name, img_crop=(128, 128)):
+        self.dataset_name = dataset_name
+        self.img_crop = img_crop
+        np.random.seed(3)
+    def load_sample_data(self, batch_size=1, is_testing=False,cal_mse=False,mse_path=''):
+        imgs_bl = []
+        imgs_sh = []
+        views=16
+        f=h5py.File(mse_path, "r")
+        label5=f['label']
+        input5=f['input']
+        for iiter,(img_bl, img_sh) in enumerate(zip(input5,label5)):
+            imgs_bl.append(img_bl)
+            imgs_sh.append(img_sh)
+        f.close()
+        imgs_bl = np.array(imgs_bl) 
+        imgs_sh = np.array(imgs_sh) 
+        return imgs_bl, imgs_sh 
+
+    def load_data(self, batch_size=1, is_testing=False,cal_mse=False,mse_path=''):
+        path = '/scratch/gilbreth/li3120/dataset/DIV2K_train_HR/Train/%s.h5' % (self.dataset_name)
+        f= h5py.File(path, "r")
+        input5=f['input']
+        ll=len(input5)
+        f.close()
+        iteration=1000
+        ipath=np.random.permutation(range(ll))
+        ipath=ipath[:1000*batch_size]
+        for i in range(iteration):
+            batch = ipath[i*batch_size:(i+1)*batch_size]
+            rd=np.random.randint(4,size=batch_size)
+            imgs_bl = []
+            imgs_sh = []
+            f= h5py.File(path, "r")
+            rdind=0
+            for img_path in batch:
+
+                label5 =f["label"]
+                img_sh=label5[img_path]
+                img_sh=np.array(img_sh)
+                if rd[rdind]==0:
+                    img_sh=np.flip(img_sh,0)
+                if rd[rdind]==1:
+                    img_sh=np.flip(img_sh,1)
+                if rd[rdind]==2:
+                    img_sh=np.rot90(img_sh,1,(0,1))
+                input5=f["input"]
+                img_bl=input5[img_path]
+                img_bl=np.array(img_bl)
+                if rd[rdind]==0:
+                    img_bl=np.flip(img_bl,0)
+                if rd[rdind]==1:
+                    img_bl=np.flip(img_bl,1)
+                if rd[rdind]==2:
+                    img_bl=np.rot90(img_bl,1,(0,1))
+                imgs_bl.append(img_bl)
+                imgs_sh.append(img_sh)
+                rdind=rdind+1
+            f.close()    
+            imgs_bl = np.array(imgs_bl) 
+            imgs_sh = np.array(imgs_sh) 
+            yield imgs_bl, imgs_sh    
+
+    def load_test_data(self, batch_size=1, is_testing=False,cal_mse=False,mse_path=''):
+        batch_images=glob('%s/*'%mse_path)
+        print(batch_images[0])    
+        imgs_bl = []
+        imgs_sh = []
+        views=16
+        for img_path in batch_images:
+            f= h5py.File(img_path, "r")
+            img_sh =f["gt"]
+            img_sh=np.expand_dims(img_sh, axis=-1)
+            img_bl=f["mvbp"]
+            img_bl=np.array(img_bl)
+            imgs_bl.append(img_bl)
+            imgs_sh.append(img_sh)
+            f.close()
+        imgs_bl = np.array(imgs_bl) / 1000.0
+        imgs_sh = np.array(imgs_sh) / 1000.0
+        return imgs_bl, imgs_sh