DBSRCNN/DBSRCNN_keras/generate_data.py

# -*- coding: utf-8 -*-
"""
Created on Thu Nov  8 22:24:55 2018

@author: fdrea
"""

import os
import cv2
import h5py
import numpy
from scipy import arange
import matplotlib.pyplot as plt
from PIL import Image    

#DATA_PATH = "./data/Trainingset/" # for train
#DATA_PATH = "./data/Train_291_aug/" # for train 291 augmentation 
#DATA_PATH = "./data/Train_291/" # for train 201 images

DATA_PATH = "./data/Test/Set5/" # for validation

#DATA_PATH = "./data/Set1/" # for try

Random_Crop = 30  # number of random patches
Patch_size = 32
label_size = 32
scale = 2

blurring_levels = 40
#b= 2



def prepare_training_data():
    names = os.listdir(DATA_PATH)
    names = sorted(names)
    nums = names.__len__()
    count = 0
    
    #data = numpy.zeros((nums * Random_Crop, 1, Patch_size, Patch_size), dtype=numpy.double)
    #label = numpy.zeros((nums * Random_Crop, 1, label_size, label_size), dtype=numpy.double)
    
     # number of saved batches in h5 file
    data = numpy.zeros((nums * Random_Crop * blurring_levels, 1, Patch_size, Patch_size), dtype=numpy.double)
    label = numpy.zeros((nums * Random_Crop * blurring_levels, 1, label_size, label_size), dtype=numpy.double)
    #imshow = plt.imshow 
    
    for i in range(nums):
        print('i = ', i)
        name = DATA_PATH + names[i]
        hr_img = cv2.imread(name, cv2.IMREAD_COLOR)
        shape = hr_img.shape

        hr_img = cv2.cvtColor(hr_img, cv2.COLOR_BGR2YCrCb)
        hr_img = hr_img[:, :, 0]

        # produce Random_Crop random coordinate to crop training img
        if(min(shape[0], shape[1]) - label_size < 0):
            continue
        Points_x = numpy.random.randint(0, min(shape[0], shape[1]) - label_size, Random_Crop)
        Points_y = numpy.random.randint(0, min(shape[0], shape[1]) - label_size, Random_Crop)
        
        for b in arange(0.1, 4.1, 0.1):
            b = round(b, 2)
            print('b = ', b)
            
            for j in range(Random_Crop):
                print('j = ', j)
                hr_patch = hr_img[Points_x[j]: Points_x[j] + label_size, Points_y[j]: Points_y[j] + label_size]
                
                #cv2.imshow("hr", hr_patch)
                
                #imshow(hr_patch)
                #plt.show()
                
                lr_patch = cv2.resize(hr_patch, (label_size // scale, label_size // scale), cv2.INTER_CUBIC)
                
                #cv2.imshow("down", lr_patch)
                
                #imshow(lr_patch)
                #plt.show()
                
                lr_patch = cv2.resize(lr_patch , (lr_patch.shape[1] * scale, lr_patch.shape[0] * scale), cv2.INTER_CUBIC)
                
                #cv2.imshow("bicubic", lr_patch)
                
                lr_patch = cv2.GaussianBlur(lr_patch, (0,0), sigmaX = b) # to blur
                
                #cv2.imshow("blur_lr", lr_patch)
                
                #imshow(lr_patch)
                #plt.show()
    
                lr_patch = lr_patch.astype(float) / 255.
                
                #imshow(lr_patch)
                #plt.show()
                
                hr_patch = hr_patch.astype(float) / 255.
                
                #imshow(hr_patch)
                #plt.show()
    
                data[count, 0, :, :] = lr_patch
                label[count, 0, :, :] = hr_patch
                #cv2.imshow("lr/255", lr_patch)
                #cv2.imshow("hr/255", hr_patch)
                #cv2.waitKey(0)
                count= count+1
                
    print('number of samples', count)        
    return data, label


def write_hdf5(data, labels, output_filename):
    """
    This function is used to save image data and its label(s) to hdf5 file.
    output_file.h5,contain data and label
    """

    x = data.astype(numpy.float32)
    y = labels.astype(numpy.float32)

    with h5py.File(output_filename, 'w') as h:
        h.create_dataset('data', data=x, shape=x.shape)
        h.create_dataset('label', data=y, shape=y.shape)
        # h.create_dataset()


def read_training_data(file):
    with h5py.File(file, 'r') as hf:
        data = numpy.array(hf.get('data'))
        label = numpy.array(hf.get('label'))
        train_data = numpy.transpose(data, (0, 2, 3, 1))
        train_label = numpy.transpose(label, (0, 2, 3, 1))
        return train_data, train_label


if __name__ == "__main__":
    data, label = prepare_training_data()
    #write_hdf5(data, label, "try_train.h5")
    
    # for training 91 images:
    #write_hdf5(data, label, "train91.h5")
    
    # for training 291 images:
    #write_hdf5(data, label, "train291.h5")
    
    # for training 291 aug images:
    #write_hdf5(data, label, "train_aug.h5")
    
    # for validation Set5:
    write_hdf5(data, label, "val.h5")
    
    #_, _a = read_training_data("train.h5")
    #_, _a = read_training_data("test.h5")