手把手教你玩转黑白图片上色【玩转华为云】

我们都知道，有很多经典的老照片，受限于那个时代的技术，只能以黑白的形式传世。尽管黑白照片别有一番风味，但是彩色照片有时候能给人更强的代入感。今天在这里为大家提供一种基于华为云ModelArts的AI技术，智能给黑白图片上色的方法，下面就紧跟小编的步骤，一步一步的来实现吧！

实验准备

华为云ModelArts平台，如对该平台不是很了解，请参考文档信息成长地图_AI开发平台ModelArts_华为云 (huaweicloud.com) ，这些文档资料可以让你完成从AI小白到AI大神的蜕变！

实验流程

一、实验环境配置

点击下方连接进入 实例感知图像上色 的 JupyterLab 页面，进入代码操作界面

https://authoring-modelarts-cnnorth4.huaweicloud.com/console/lab?share-url-b64=aHR0cHM6Ly9jbm5vcnRoNC1tb2RlbGh1Yi1tb2RlbHMub2JzLmNuLW5vcnRoLTQubXlod2Nsb3Vkcy5jb20vaXRlbS83OTJjNzlhNS1jYzI5LTRhNjUtYmUzNy01MTRhYjk1OWE2OTIvMS4wLjAvSW5zdENvbG9yaXphdGlvbi5pcHluYg%3D%3D&galleryitemid=792c79a5-cc29-4a65-be37-514ab959a692

不要使用默认的配置信息，这里我们选择GPU: 1*V100|CPU: 8核 64GB的环境，默认的信息在配置 Detectron2时，会报错。

切换流程见下图：

切换完成后，点击确定

2、 实验步骤

2.1下载代码和数据

import os
!wget https://obs-aigallery-zc.obs.cn-north-4.myhuaweicloud.com/clf/code/InstColorization.zip
os.system('unzip InstColorization.zip')

选择代码后，按crtl + enter 键运行，或者点击工具栏的开始按钮

运行成功后，[*]会变成对应的[1]值，或者工具栏左下方的 Busy 变成 Idle 

2.2安装依赖库

!gcc --version

!pip install torch==1.5 torchvision==0.6
!pip install cython pyyaml==5.1
!pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
!pip install dominate==2.4.0
!pip install detectron2==0.1.3 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/index.html
!pip install opencv-python
!pip install scikit-image
    
import torch, torchvision
print(torch.__version__, torch.cuda.is_available())

2.3准备上色

2.3.1切换目录

cd InstColorization/

2.3.2配置 Detectron2

from os.path import join, isfile, isdir
from os import listdir
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
from argparse import ArgumentParser

import detectron2
from detectron2.utils.logger import setup_logger
setup_logger()

import numpy as np
import cv2

# import some common detectron2 utilities
from detectron2 import model_zoo
from detectron2.engine import DefaultPredictor
from detectron2.config import get_cfg

import torch

cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file("COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x.yaml"))
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.7
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x.yaml")
predictor = DefaultPredictor(cfg)

input_dir = "example"
image_list = [f for f in listdir(input_dir) if isfile(join(input_dir, f))]
output_npz_dir = "{0}_bbox".format(input_dir)
if os.path.isdir(output_npz_dir) is False:
    print('Create path: {0}'.format(output_npz_dir))
    os.makedirs(output_npz_dir)

for image_path in image_list:
    img = cv2.imread(join(input_dir, image_path))
    lab_image = cv2.cvtColor(img, cv2.COLOR_BGR2LAB)
    l_channel, a_channel, b_channel = cv2.split(lab_image)
    l_stack = np.stack([l_channel, l_channel, l_channel], axis=2)
    outputs = predictor(l_stack)
    save_path = join(output_npz_dir, image_path.split('.')[0])
    pred_bbox = outputs["instances"].pred_boxes.to(torch.device('cpu')).tensor.numpy()
    pred_scores = outputs["instances"].scores.cpu().data.numpy()
    np.savez(save_path, bbox = pred_bbox, scores = pred_scores)

!ls example_bbox

2.4图像上色

import sys
import time
from options.train_options import TestOptions
from models import create_model

import torch
from tqdm import tqdm_notebook

from fusion_dataset import Fusion_Testing_Dataset
from util import util
import multiprocessing
multiprocessing.set_start_method('spawn', True)

torch.backends.cudnn.benchmark = True

sys.argv = [sys.argv[0]]
opt = TestOptions().parse()

save_img_path = opt.results_img_dir
if os.path.isdir(save_img_path) is False:
    print('Create path: {0}'.format(save_img_path))
    os.makedirs(save_img_path)
opt.batch_size = 1
dataset = Fusion_Testing_Dataset(opt, -1)
dataset_loader = torch.utils.data.DataLoader(dataset, batch_size=opt.batch_size)

dataset_size = len(dataset)
print('#Testing images = %d' % dataset_size)

model = create_model(opt)
model.setup_to_test('coco_finetuned_mask_256_ffs')

count_empty = 0
for data_raw in tqdm_notebook(dataset_loader):
    data_raw['full_img'][0] = data_raw['full_img'][0].cuda()
    if data_raw['empty_box'][0] == 0:
        data_raw['cropped_img'][0] = data_raw['cropped_img'][0].cuda()
        box_info = data_raw['box_info'][0]
        box_info_2x = data_raw['box_info_2x'][0]
        box_info_4x = data_raw['box_info_4x'][0]
        box_info_8x = data_raw['box_info_8x'][0]
        cropped_data = util.get_colorization_data(data_raw['cropped_img'], opt, ab_thresh=0, p=opt.sample_p)
        full_img_data = util.get_colorization_data(data_raw['full_img'], opt, ab_thresh=0, p=opt.sample_p)
        model.set_input(cropped_data)
        model.set_fusion_input(full_img_data, [box_info, box_info_2x, box_info_4x, box_info_8x])
        model.forward()
    else:
        count_empty += 1
        full_img_data = util.get_colorization_data(data_raw['full_img'], opt, ab_thresh=0, p=opt.sample_p)
        model.set_forward_without_box(full_img_data)
    model.save_current_imgs(join(save_img_path, data_raw['file_id'][0] + '.png'))
print('{0} images without bounding boxes'.format(count_empty))

2.5展示上色结果

def imshow(img):
    import IPython
    import cv2
    _, ret = cv2.imencode('.jpg', img) 
    i = IPython.display.Image(data=ret)
    IPython.display.display(i)

img_name_list = ['000000022969', '000000023781', '000000046872', '000000050145']
# 修改对应索引使用不同图片 0-3
show_index = 1

img = cv2.imread('example/'+img_name_list[show_index]+'.jpg')
lab_image = cv2.cvtColor(img, cv2.COLOR_BGR2LAB)
l_channel, _, _ = cv2.split(lab_image)

img = cv2.imread('results/'+img_name_list[show_index]+'.png')
lab_image = cv2.cvtColor(img, cv2.COLOR_BGR2LAB)
_, a_pred, b_pred = cv2.split(lab_image)
a_pred = cv2.resize(a_pred, (l_channel.shape[1], l_channel.shape[0]))
b_pred = cv2.resize(b_pred, (l_channel.shape[1], l_channel.shape[0]))
gray_color = np.ones_like(a_pred) * 128

gray_image = cv2.cvtColor(np.stack([l_channel, gray_color, gray_color], 2), cv2.COLOR_LAB2BGR)
color_image = cv2.cvtColor(np.stack([l_channel, a_pred, b_pred], 2), cv2.COLOR_LAB2BGR)

# save_img_path = 'results_origin/'
# if os.path.isdir(save_img_path) is False:
#     print('Create path: {0}'.format(save_img_path))
#     os.makedirs(save_img_path)

# cv2.imwrite('results_origin/'+img_name_list[show_index]+'.png', color_image)

imshow(np.concatenate([gray_image, color_image], 1))

到此为止，一个官方的示例，就算完成了！

如果想要给其他的黑白图片上色，又该怎么完成呢？继续往下看，嘿嘿

在2.5步骤中，有这样几句代码，通过对比可以发现，原始图集在example文件夹下，通过切换下标，可以给不同的图片完成上色操作，

img_name_list = ['000000022969', '000000023781', '000000046872', '000000050145']
# 修改对应索引使用不同图片 0-3
show_index = 1

img = cv2.imread('example/'+img_name_list[show_index]+'.jpg')
lab_image = cv2.cvtColor(img, cv2.COLOR_BGR2LAB)
l_channel, _, _ = cv2.split(lab_image)

img = cv2.imread('results/'+img_name_list[show_index]+'.png')
lab_image = cv2.cvtColor(img, cv2.COLOR_BGR2LAB)
_, a_pred, b_pred = cv2.split(lab_image)

将本地图片，通过菜单栏的upload files按钮，上传到example文件夹下

重复步骤2.3,2.4,2.5，其中2.5中需要修改照片的信息，将照片123加入数组中，同时指定上色图片索引为4

def imshow(img):
    import IPython
    import cv2
    _, ret = cv2.imencode('.jpg', img) 
    i = IPython.display.Image(data=ret)
    IPython.display.display(i)

img_name_list = ['000000022969', '000000023781', '000000046872', '000000050145', '123']
# 修改对应索引使用不同图片 0-3
show_index = 4

img = cv2.imread('example/'+img_name_list[show_index]+'.jpg')
lab_image = cv2.cvtColor(img, cv2.COLOR_BGR2LAB)
l_channel, _, _ = cv2.split(lab_image)

img = cv2.imread('results/'+img_name_list[show_index]+'.png')
lab_image = cv2.cvtColor(img, cv2.COLOR_BGR2LAB)
_, a_pred, b_pred = cv2.split(lab_image)
a_pred = cv2.resize(a_pred, (l_channel.shape[1], l_channel.shape[0]))
b_pred = cv2.resize(b_pred, (l_channel.shape[1], l_channel.shape[0]))
gray_color = np.ones_like(a_pred) * 128

gray_image = cv2.cvtColor(np.stack([l_channel, gray_color, gray_color], 2), cv2.COLOR_LAB2BGR)
color_image = cv2.cvtColor(np.stack([l_channel, a_pred, b_pred], 2), cv2.COLOR_LAB2BGR)

# save_img_path = 'results_origin/'
# if os.path.isdir(save_img_path) is False:
#     print('Create path: {0}'.format(save_img_path))
#     os.makedirs(save_img_path)

# cv2.imwrite('results_origin/'+img_name_list[show_index]+'.png', color_image)

imshow(np.concatenate([gray_image, color_image], 1))

就能完整展示自己的上色图片啦