python常用代码块 | python 人脸识别（腾讯 gfpgan方案）

python 人脸识别（腾讯 gfpgan方案）

实现功能

使用腾讯 gfpgan方案进行人脸识别并高清修复

1.首先在github下载源文件https://github.com/TencentARC/GFPGAN

左边是从视频中截取的人脸，右边是修复后的结果，只能说这效果杠杠的

2.代码修改后，运行时会在’xxx\gfpgan\weights’文件夹中自动下载GFPGANv1.3.pth
模型文件，同时在程序的文件夹’xxx\notepad\gfpgan\weights’下会下载’detection_Resnet50_Final.pth’， ‘parsing_parsenet.pth’两个模型，
可以自动运行时自动下载相应模型到对应位置，
也可以在如下链接’ ’下载这三个模型，然后放到对应的文件夹下

核心代码分三步：

1)定义关键参数(未做修改)
upscale = 2
arch ='clean'
channel_multiplier=2
bg_upsampler = None
aligned = False
only_center_face = False
weight=0.5

2)调用GFPGANer(...)
restorer = GFPGANer(
    model_path=model_path,
    upscale=upscale,
    arch=arch,
    channel_multiplier=channel_multiplier,
    bg_upsampler=bg_upsampler)

3)
cropped_faces, restored_faces, restored_img = restorer.enhance(
            input_img,
            has_aligned=aligned,
            only_center_face=only_center_face,
            paste_back=True,
            weight=weight)
---restored_faces即是我们需要的图片数据

完整代码如下：

from gfpgan import GFPGANer
import torch,os,glob,cv2,sys
import numpy as np
#from basicsr.utils import imwrite

baseDirPath = sys.path[0]
print(baseDirPath)

###################################################################
#读含中文/韩文/日文等特殊字符路径的图片
def cv_imread(in_path):
  im = cv2.imdecode(np.fromfile(in_path, dtype=np.uint8),-1)
  return im
  
#路径中有中文名 cv2写
def cv_imwrite(out_path, imp_np):
  imp_type = '.' + out_path.split('.')[-1]
  cv2.imencode(imp_type, imp_np)[1].tofile(out_path)
###################################################################

#模型位置
model_path = baseDirPath + '\\gfpgan\\weights\\GFPGANv1.3.pth'

#如下参数固定不变
upscale = 2
arch ='clean'
channel_multiplier=2
bg_upsampler = None
aligned = False
only_center_face = False
weight=0.5

if not torch.cuda.is_available():  # CPU
    bg_upsampler = None

restorer = GFPGANer(
    model_path=model_path,
    upscale=upscale,
    arch=arch,
    channel_multiplier=channel_multiplier,
    bg_upsampler=bg_upsampler)

input_path = baseDirPath + '\\pic\\src'
output_path = baseDirPath + '\\pic\\dst'

if os.path.isfile(input_path):
    img_list = [input_path]
else:
    img_list = sorted(glob.glob(os.path.join(input_path, '*')))
print(img_list)

for img_path in img_list:
    img_name = os.path.basename(img_path)
    
    print(f'processing {img_name} ...')
    basename, ext = os.path.splitext(img_name)
    
    input_img = cv_imread(img_path)
    
    cropped_faces, restored_faces, restored_img = restorer.enhance(
            input_img,
            has_aligned=aligned,
            only_center_face=only_center_face,
            paste_back=True,
            weight=weight)
    
    # save faces
    for idx, restored_face in enumerate(restored_faces):
        # save restored face
        save_restore_path = os.path.join(output_path,  f'{basename}_{idx:02d}.png')
        cv_imwrite(save_restore_path, restored_face)

print(f'Results are in the [{output_path}] folder.')

备注：如果输入图片尺寸较小或质量较差，会出现很多奇葩结果，如这样：

源码：https://download.csdn.net/download/mjc1321/89070575