スポンサーリンク
This time I tried RealESRGAN on google colab, which allows for higher resolution.
This article is a reminder to get RealESRGAN working.
contents
スポンサーリンク
abstract
How to use RealESRGAN on google colab to enhance images
1.requirement
Google Colab
git:clone https://github.com/xinntao/Real-ESRGAN.git
model : RealESRGAN_x4plus
2.Environment construction & code writing
2.1 install＆download
Install the necessary modules by cloning from git with the following command.
and then download the weights to be used in the model.
!git clone https://github.com/xinntao/Real-ESRGAN.git %cd Real-ESRGAN !apt-get install ffmpeg !pip install pydub !pip install basicsr !pip install facexlib !pip install gfpgan !pip install -r requirements.txt !python setup.py develop !wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P experiments/pretrained_models
2.2 Model Execution ~ Result Saving
Describes the process of loading the specified model and arguments and saving them as high quality images.
import cv2
import glob
import os
from basicsr.archs.rrdbnet_arch import RRDBNet
from realesrgan import RealESRGANer
import subprocess
from pydub import AudioSegment
from google.colab.patches import cv2_imshow
model_path = "experiments/pretrained_models/RealESRGAN_x4plus.pth" #@param {type:"string"}
path = 'path/to/image/' #@param {type:"string"}
# variable
block = 23
netscale = 4
tile = 0
tile_pad = 10
pre_pad = 0
half = False
face_enhance = False
outscale = 8 #@param {type:"number"}
output_dir = '/path/to/output_dir' #@param {type:"string"}
resize_rate = 0.5
limit_img_size = 1000
if 'RealESRGAN_x4plus_anime_6B.pth' in model_path:
block = 6
elif 'RealESRGAN_x2plus.pth' in model_path:
netscale = 2
model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=block, num_grow_ch=32, scale=netscale)
upsampler = RealESRGANer(
scale=netscale,
model_path=model_path,
model=model,
tile=tile,
tile_pad=tile_pad,
pre_pad=pre_pad,
half=half)
if face_enhance:
from gfpgan import GFPGANer
face_enhancer = GFPGANer(
model_path='path/to/model',
upscale=outscale,
arch='clean',
channel_multiplier=2,
bg_upsampler=upsampler)
else:
pass
os.makedirs(output_dir, exist_ok=True)
img = cv2.imread(path)
height = img.shape[0]
width = img.shape[1]
print('original image:', height, width)
if height < width:
long_size = width
else:
long_size = height
# Resize the input image if its size is larger than 1000pixel, because it will exceed the memory.
if limit_img_size < long_size:
resize_rate = limit_img_size/long_size
in_img = cv2.resize(img , (int(width*resize_rate), int(height*resize_rate)))
print('resize image', in_img.shape[0], in_img.shape[1])
else:
in_img = img
if face_enhance:
# _, _, output = face_enhancer.enhance(img, has_aligned=False, only_center_face=False, paste_back=True)
pass
else:
output, _ = upsampler.enhance(in_img, outscale=outscale)
save_path = path/to/save
cv2.imwrite(save_path, output)
del output
del in_img
del img
del upsampler
3.result
original image
Zoom in for comparison.
left：normal / right：After high image quality
The skin and eye area are more beautiful in the image after the higher resolution.
The hair was a little crushed and needs to be adjusted
