utils.py

import tensorflow as tf
import tensorlayer as tl
from tensorlayer.prepro import *
from config import config, log_config
from skimage import feature
from skimage import color
from scipy.ndimage.filters import gaussian_filter

import scipy
import numpy as np
import cv2
import math
import random

import os
import fnmatch

def read_all_imgs(file_name_list, path = '', mode = 'RGB'):
    imgs = []
    for idx in range(0, len(file_name_list)):
        imgs.append(get_images(file_name_list[idx], path, mode))

    return imgs

def get_images(file_name, path, mode):
    """ Input an image path and name, return an image array """
    # return scipy.misc.imread(path + file_name).astype(np.float)
    if mode is 'RGB':
        image = (scipy.misc.imread(path + file_name, mode='RGB')/255.).astype(np.float32)
    elif mode is 'GRAY':
        image = (scipy.misc.imread(path + file_name, mode='P')/255.).astype(np.float32)
        image = np.expand_dims(image, axis = 2)
    elif mode is 'NPY':
        image = np.load(path + file_name)
        image = image / 3.275
        image = np.expand_dims(image, axis = 2)
    elif mode is 'DEPTH':
        image = (np.float32(cv2.imread(path + file_name, cv2.IMREAD_UNCHANGED))/10.)[:, :, 1]
        ## If you train the network with the SYNDOF dataset (this is the original SYNDOF dataset) shared in this repository.
        ## The SYNDOF datasets's maximum COC value is 15 and we saved the defocus map with the COC value.
        ## (The paper say that maximum COC value is 28, becuase the blur kernel of orignal SYNDOF dataset visually had the maximaum coc value of 28 when it was generated with max_coc=15.)

        image = image / 15

        ## If you train the network with the new SYNDOF dataset generated with the codes in "https://github.com/codeslake/SYNDOF".
        ## We save the sigma value (max=7) in the code, where
        ## sigma = (max_coc-1)/4, when max_coc = 29, max_sigma = 7

        # image = image / 7

        image = np.expand_dims(image, axis = 2)

    return image

def t_or_f(arg):
    ua = str(arg).upper()
    if 'TRUE'.startswith(ua):
       return True
    elif 'FALSE'.startswith(ua):
       return False
    else:
       pass

def _tf_fspecial_gauss(size, sigma):
    """Function to mimic the 'fspecial' gaussian MATLAB function
    """
    x_data, y_data = np.mgrid[-size//2 + 1:size//2 + 1, -size//2 + 1:size//2 + 1]

    x_data = np.expand_dims(x_data, axis=-1)
    x_data = np.expand_dims(x_data, axis=-1)

    y_data = np.expand_dims(y_data, axis=-1)
    y_data = np.expand_dims(y_data, axis=-1)

    x = tf.constant(x_data, dtype=tf.float32)
    y = tf.constant(y_data, dtype=tf.float32)

    g = tf.exp(-((x**2 + y**2)/(2.0*sigma**2)))
    return g / tf.reduce_sum(g)

def refine_image(img):
    h, w = img.shape[:2]

    return img[0 : h - h % 16, 0 : w - w % 16]

def random_crop(images, resize_shape, is_gaussian_noise = False):
    images_list = None
    h, w = resize_shape[:2]
    max_size_limit = 800

    for i in np.arange(len(images)):
        image = np.copy(images[i])
        shape = np.array(image.shape[:2])

        if shape.min() <= h:
            ratio = resize_shape[shape.argmin()]/float(shape.min())
            resize_w = int(math.floor(shape[1] * ratio)) + 1
            resize_h = int(math.floor(shape[0] * ratio)) + 1
            image = cv2.resize(image, (resize_w, resize_h))

        if shape.min() > max_size_limit:
            ratio = max_size_limit/float(shape.min())
            resize_w = int(math.floor(shape[1] * ratio)) + 1
            resize_h = int(math.floor(shape[0] * ratio)) + 1
            image = cv2.resize(image, (resize_w, resize_h))

        if is_gaussian_noise:
            image = add_gaussian_noise(image)

        cropped_image = tl.prepro.crop(image, wrg=w, hrg=h, is_random=True)
        augmented_image = _random_flip(cropped_image)
        angles = np.array([1, 2, 3, 4])
        angle = np.random.choice(angles)
        augmented_image = _random_rotation(augmented_image, angle)
        image = np.expand_dims(augmented_image, axis=0)

        images_list = np.copy(image) if i == 0 else np.concatenate((images_list, image), axis = 0)

    return images_list

def crop_pair_with_different_shape_images(images, labels, resize_shape, is_gaussian_noise = False):
    images_list = None
    labels_list = None
    h, w = resize_shape[:2]
    max_size_limit = 800

    for i in np.arange(len(images)):
        image = np.copy(images[i])
        label = np.copy(labels[i])
        shape = np.array(image.shape[:2])

        if shape.min() <= h:
            ratio = resize_shape[shape.argmin()]/float(shape.min())
            resize_w = int(math.floor(shape[1] * ratio)) + 1
            resize_h = int(math.floor(shape[0] * ratio)) + 1
            image = cv2.resize(image, (resize_w, resize_h))
            label = np.expand_dims(cv2.resize(label[:, :, 0], (resize_w, resize_h)), axis = 2)

        if shape.min() > max_size_limit:
            ratio = max_size_limit/float(shape.min())
            resize_w = int(math.floor(shape[1] * ratio)) + 1
            resize_h = int(math.floor(shape[0] * ratio)) + 1
            image = cv2.resize(image, (resize_w, resize_h))
            label = np.expand_dims(cv2.resize(label[:, :, 0], (resize_w, resize_h)), axis = 2)

        if is_gaussian_noise:
            image = add_gaussian_noise(image)

        concatenated_images = np.concatenate((image, label), axis = 2)
        cropped_images = tl.prepro.crop(concatenated_images, wrg=w, hrg=h, is_random=True)
        augmented_images = _random_flip(cropped_images)
        angles = np.array([1, 2, 3, 4])
        angle = np.random.choice(angles)
        augmented_images = _random_rotation(augmented_images, angle)

        image = np.expand_dims(augmented_images[:, :, 0:3], axis=0)
        label = np.expand_dims(np.expand_dims(augmented_images[:, :, 3], axis=3), axis=0)

        images_list = np.copy(image) if i == 0 else np.concatenate((images_list, image), axis = 0)
        labels_list = np.copy(label) if i == 0 else np.concatenate((labels_list, label), axis = 0)

    return images_list, labels_list

def add_gaussian_noise(image):
    image = image.astype(np.float32)
    shape = image.shape[:2]

    mean = 0
    var = random.uniform(0,0.1)
    sigma = var ** 0.5
    gamma = 0.25
    alpha = 0.75
    beta = 1 - alpha

    gaussian = np.random.normal(loc=mean, scale = sigma, size = (shape[0], shape[1], 1)).astype(np.float32)
    gaussian = np.concatenate((gaussian, gaussian, gaussian), axis = 2)
    #gaussian_img = image * 0.75 + 0.25 * gaussian + 0.25
    gaussian_img = cv2.addWeighted(image, alpha, beta * gaussian, beta, gamma)

    return gaussian_img

    # noise_sigma = 0.01
    # h = image.shape[0]
    # w = image.shape[1]
    # noise = np.random.randn(h, w) * noise_sigma

    # noisy_image = np.zeros(image.shape, np.float64)
    # if len(image.shape) == 2:
    #     noisy_image = image + noise
    # else:
    #     noisy_image[:,:,0] = image[:,:,0] + noise
    #     noisy_image[:,:,1] = image[:,:,1] + noise
    #     noisy_image[:,:,2] = image[:,:,2] + noise

    # """
    # print('min,max = ', np.min(noisy_image), np.max(noisy_image))
    # print('type = ', type(noisy_image[0][0][0]))
    # """

    # return noisy_image

def _random_flip(images):
    flipped_images = tl.prepro.flip_axis(images, axis=0, is_random=True)

    return flipped_images

def _random_rotation(images, angle):
    if angle != 4:
        rotated_images = np.rot90(images, angle)
    else:
        rotated_images = images

    return rotated_images

def _get_file_path(path, regex):
    file_path = []
    for root, dirnames, filenames in os.walk(path):
        for i in np.arange(len(regex)):
            for filename in fnmatch.filter(filenames, regex[i]):
                file_path.append(os.path.join(root, filename))

    return file_path

def remove_file_end_with(path, regex):
    file_paths = _get_file_path(path, [regex])

    for i in np.arange(len(file_paths)):
        os.remove(file_paths[i])

def save_images(images, size, image_path='_temp.png'):
    if len(images.shape) == 3:  # Greyscale [batch, h, w] --> [batch, h, w, 1]
        images = images[:, :, :, np.newaxis]

    def merge(images, size):
        h, w = images.shape[1], images.shape[2]
        img = np.zeros((h * size[0], w * size[1], 3))
        for idx, image in enumerate(images):
            i = idx % size[1]
            j = idx // size[1]
            img[j * h:j * h + h, i * w:i * w + w, :] = image
        return img

    def imsave(images, size, path):
        return scipy.misc.toimage(merge(images, size), cmin = 0., cmax = 1.).save(path)

    assert len(images) <= size[0] * size[1], "number of images should be equal or less than size[0] * size[1] {}".format(len(images))

    return imsave(images, size, image_path)

def fix_image_tf(image, norm_value):
    return tf.cast(image / norm_value * 255., tf.uint8)

def norm_image_tf(image):
    image = image - tf.reduce_min(image, axis = [1, 2, 3], keepdims=True)
    image = image / tf.reduce_max(image, axis = [1, 2, 3], keepdims=True)
    return tf.cast(image * 255., tf.uint8)

def norm_image(image, axis = (1, 2, 3)):
    image = image - np.amin(image, axis = axis, keepdims=True)
    image = image / np.amax(image, axis = axis, keepdims=True)
    return image

def get_disc_accuracy(logits, labels):
    acc = 0.
    for i in np.arange(len(logits)):
        tp = 0
        logits[i] = np.round(np.squeeze(logits[i])).astype(int)
        temp = logits[i]
        tp = tp + len(temp[np.where(temp == labels[i])])
        acc = acc + (tp / float(len(logits[i])))
    return acc / float(len(labels))