Python给图像制作Mosaic

    最近研究超分辨率过程中，脑洞大开，考虑到对于马赛克图像的恢复实验，说GAN就GAN，于是制作了一批马赛克dataset，打算利用SRGAN网络结构对其进行复原。

只是实验性的过程，目前还不知道效果。

先送上马赛克处理过程，这里参考来Ps打马算法的实现。

import scipy.misc as misc
import os,sys
import random
import numpy as np
from multiprocessing import Pool
import shutil

MOSAIC_RANGE = [1, 3]

input_dir = '../data/Mosaic_LR/'
output_dir = '../data/Mosaic_LR_msk/'
if not os.path.exists(input_dir):
    os.mkdir(input_dir)
if not os.path.exists(output_dir):
    os.mkdir(output_dir)

index = 0
# 打码函数
def drawMask(file_name):
    #print(file_name)
    global index
    index += 1

    try:
        sys.stdout.write('\r>>Converting image %d/100000 ' % (index))
        sys.stdout.flush()
        img = misc.imread(file_name).astype(dtype=np.float)
        img_out = img.copy()
        row, col, channel = img.shape
        half_patch = np.random.randint(MOSAIC_RANGE[0], MOSAIC_RANGE[1]+1, 1)[0]
        for i in range(half_patch, row - 1 - half_patch, half_patch):
            for j in range(half_patch, col - 1 - half_patch, half_patch):
                k1 = random.random() - 0.5
                k2 = random.random() - 0.5
                m = np.floor(k1 * (half_patch * 2 + 1))
                n = np.floor(k2 * (half_patch * 2 + 1))
                h = int((i + m) % row)
                w = int((j + n) % col)
                img_out[i - half_patch:i + half_patch, j - half_patch:j + half_patch, :] = img[h, w, :]

        save_name = os.path.join(output_dir, file_name.split('/')[-1])
        misc.imsave(save_name,img_out)
    except IOError as e:
        print('could not read:',jpg)
        print('error:',e)
        print('skip it\n')


if __name__ == "__main__":
    dataset_dir = input_dir
    image_list = os.listdir(dataset_dir)
    image_list = [os.path.join(dataset_dir, _) for _ in image_list]
    image_list = image_list[:100000]

    p = Pool(5)
    p.map(drawMask, image_list)


    sys.stdout.write('Convert Over!\n')
    sys.stdout.flush()

原图是这样的：

打马后：

原图：

打Ma:

原图：

轻微打ma:

当然也可以局部打ma，考虑到训练数据集的需要，这里我都是对整张图打马处理的。

升级版：

主要加了异常处理，比如图像读取错误，图像本身错误损坏等导致图像异常，造成程序终止，处理方式：一旦发现异常，自动清理图片。

import scipy.misc as misc
import os,sys
import random
import numpy as np
from multiprocessing import Pool
import shutil

MOSAIC_RANGE = [2, 7]

input_dir = '../datasets/mosaic/trainB_All/'
output_dir = '../datasets/mosaic/trainA/'
move_dir = '../datasets/mosaic/trainB/'

if not os.path.exists(input_dir):
    os.mkdir(input_dir)
if not os.path.exists(output_dir):
    os.mkdir(output_dir)

if not os.path.exists(move_dir):
    os.mkdir(move_dir)
index = 0

# 打码函数
def drawMask(file_name):
    #print(file_name)
    global index
    index += 1

    try:
        sys.stdout.write('\r>>Converting image %d/20000 ' % (index))
        sys.stdout.flush()
        img = misc.imread(file_name).astype(dtype=np.float)
        img_out = img.copy()
        row, col, channel = img.shape
        half_patch = np.random.randint(MOSAIC_RANGE[0], MOSAIC_RANGE[1]+1, 1)[0]
        for i in range(half_patch, row - 1 - half_patch, half_patch):
            for j in range(half_patch, col - 1 - half_patch, half_patch):
                k1 = random.random() - 0.5
                k2 = random.random() - 0.5
                m = np.floor(k1 * (half_patch * 2 + 1))
                n = np.floor(k2 * (half_patch * 2 + 1))
                h = int((i + m) % row)
                w = int((j + n) % col)
                img_out[i - half_patch:i + half_patch, j - half_patch:j + half_patch, :] = img[h, w, :]

        save_name = os.path.join(output_dir, file_name.split('/')[-1])
        misc.imsave(save_name,img_out)
    except IOError as e:
        print('could not read:',file_name)
        print('error:',e)
        print('skip it\n')
        shutil.move(file_name, '/home/gavin/Desktop/face_test/')
    except TypeError as err:
        print('TypeError:', file_name)
        print('error:', err)
        print('skip it\n')
        shutil.move(file_name, '/home/gavin/Desktop/valueErr/')
    else:
        shutil.move(file_name, move_dir)


if __name__ == "__main__":
    dataset_dir = input_dir
    image_list = os.listdir(dataset_dir)
    image_list = [os.path.join(dataset_dir, _) for _ in image_list]
    image_list = image_list[:3000]

    p = Pool(5)
    p.map(drawMask, image_list)


    sys.stdout.write('Convert Over!\n')
    sys.stdout.flush()