def main():
import sys
args = get_FaceSR_opt()
img_path = args.image
img = utils.read_cv2_img(img_path)
sr_model = SRGANModel(args, is_train=False)
sr_model.load()
def sr_forward(img, padding=0.5, moving=0.1):
img_aligned, M = dlib_detect_face(img,
padding=padding,
image_size=(128, 128),
moving=moving)
input_img = torch.unsqueeze(_transform(Image.fromarray(img_aligned)),
0)
sr_model.var_L = input_img.to(sr_model.device)
sr_model.test()
output_img = sr_model.fake_H.squeeze(0).cpu().numpy()
output_img = np.clip((np.transpose(output_img,
(1, 2, 0)) / 2.0 + 0.5) * 255.0, 0,
255).astype(np.uint8)
rec_img = face_recover(output_img, M * 4, img)
return output_img, rec_img
output_img, rec_img = sr_forward(img)
utils.save_image(output_img, 'output_face.jpg')
utils.save_image(rec_img, 'output_img.jpg')
def generate_img_dir(img_dir, out_dir):
"""
处理视频文件夹，输出至文件夹
"""
# img_dir = os.path.join(DATA_DIR, 'imgs')
# out_dir = os.path.join(DATA_DIR, 'outs')
mkdir_if_not_exist(out_dir)
paths_list, names_list = traverse_dir_files(img_dir)
ig = ImgGenerator()
count = 0
for path, name in zip(paths_list, names_list):
print('[Info] 图像路径: {}'.format(path))
# img_path = os.path.join(DATA_DIR, 'imgs', 'input_4.jpg')
sr_name = name.split('.')[0] + '.jpg'
out_path = os.path.join(out_dir, sr_name)
try:
img = utils.read_cv2_img(path)
rec_img = ig.sr_forward(img=img) # 超分辨率图像
utils.save_image(rec_img, out_path)
except Exception as e:
print('[Info] 异常: {}'.format(e))
print('[Info] 图像: {}'.format(path))
count += 1
if count % 100 == 0:
print('[Info] count {}'.format(count))
print('[Info] 视频文件处理完成!')
print('[Info] 输出文件夹: {}'.format(out_dir))
def main():
# ---------------------------Test ---------------------------------
print("Start SR test")
img = utils.read_cv2_img(i)
#img = img.resize((128,128))
in_img = torch.unsqueeze(_transform(Image.fromarray(img)), 0)
sr_model.var_L = in_img.to(sr_model.device)
sr_model.test()
#visuals = sr_model.fake_H.squeeze(0).cpu().numpy()
visuals = sr_model.fake_H.detach().float().cpu()
image_numpy = utils.tensor2im(visuals, show_size=317)
image_numpy = np.reshape(image_numpy, (-1, 317, 3))
image_numpy = cv2.resize(image_numpy, (img.shape[0], img.shape[1]))
print('End test')
print()
# ----------------------------------End test--------------------------
# -----------------------------SR img combine Original img --------------------------------------------
#row,cols, channels = crop_img.shape
start_x = result[i][j]['box'][0]
start_y = result[i][j]['box'][1]
end_x = result[i][j]['box'][0] + result[i][j]['box'][2]
end_y = result[i][j]['box'][1] + result[i][j]['box'][3]
area = (start_x, start_y, end_x, end_y)
crop_img = crop_img.resize((int(end_x - start_x), int(end_y - start_y)))
px = f_img.load()
c_px = crop_img.load()
c_x = 0
c_x_max = crop_img.width
c_y = 0
c_y_max = crop_img.height
# print(crop_img.width,crop_img.height)
# print(end_x-start_x,end_y-start_y)
for q in range(start_x, end_x):
c_y = 0
for k in range(start_y, end_y):
# print(c_px[c_x,c_y])
try:
px[q, k] = c_px[c_x, c_y]
if (c_y < c_y_max):
c_y = c_y + 1
except:
if (c_y < c_y_max):
c_y = c_y + 1
pass
if (c_x < c_x_max):
c_x = c_x + 1
my_count = my_count + 1
# f_img.save('/content/drive/MyDrive/Face-Super-Resolution/Video_test/final/{}.png'.format(i.split('/')[-1].split('.')[0]))
f_img.save(args.final_path +
'/{}.png'.format(i.split('/')[-1].split('.')[0]))
r_c = r_c + 1
# cv2.imwrite("/content/drive/MyDrive/PARK/IP-FSRGAN/final/{}.png".format(r_c),f_img)
print("End SR img combine to ori img")
print()
# -----------------------------SR img combine Original img End --------------------------------------------
end_time = time.time()
print("time = {}".format(end_time - start_time))
def generate_img(img_path, out_path):
print('[Info] 图像路径: {}'.format(img_path))
# 图像
img = utils.read_cv2_img(img_path)
img_name = img_path.split('/')[-1].split('.')[0]
# 输出文件夹
outs_dir = os.path.join(DATA_DIR, 'outs')
mkdir_if_not_exist(outs_dir)
ig = ImgGenerator()
rec_img = ig.sr_forward(img=img) # 超分辨率图像
# 输出图像
# out_path = os.path.join(outs_dir, '{}.output.2.jpg'.format(img_name))
utils.save_image(rec_img, out_path)
parser.add_argument('--D_nf', type=int, default=64)
# data dir
parser.add_argument('--pretrain_model_G', type=str, default='90000_G.pth')
parser.add_argument('--pretrain_model_D', type=str, default=None)
args = parser.parse_args()
return args
sr_model = SRGANModel(get_FaceSR_opt(), is_train=False)
sr_model.load()
def sr_forward(img, padding=0.5, moving=0.1):
img_aligned, M = dlib_detect_face(img, padding=padding, image_size=(128, 128), moving=moving)
input_img = torch.unsqueeze(_transform(Image.fromarray(img_aligned)), 0)
sr_model.var_L = input_img.to(sr_model.device)
sr_model.test()
output_img = sr_model.fake_H.squeeze(0).cpu().numpy()
output_img = np.clip((np.transpose(output_img, (1, 2, 0)) / 2.0 + 0.5) * 255.0, 0, 255).astype(np.uint8)
rec_img = face_recover(output_img, M * 4, img)
return output_img, rec_img
img_path = 'input.jpg'
img = utils.read_cv2_img(img_path)
output_img, rec_img = sr_forward(img)
utils.save_image(output_img, 'output_face.jpg')
utils.save_image(rec_img, 'output_img.jpg')
sr_model.test()
output_img = sr_model.fake_H.squeeze(0).cpu().numpy()
output_img = np.clip((np.transpose(output_img,
(1, 2, 0)) / 2.0 + 0.5) * 255.0, 0,
255).astype(np.uint8)
rec_img = face_recover(output_img, M * 4, img)
return output_img, rec_img
#cut으로 나눴을 때, 폴더안에 들어가있는 14장의 디렉토리 (단, 14장 따로 있어서
# 이를 포괄하는 디렉토리를 지정해줘야함.)
# 그리고, 화질이 안좋은 사진 14개는 change_image의 하위폴더인 original에 담겨있어야함
# 또, 이미지 화질 보정을 통한 사진이 저장될 경로로 빈폴더 resolution도 존재해야한다.
img_path = r'C:\Users\YongTaek\Desktop\change_image'
origin_path = os.path.join(img_path, 'original')
list_photo = os.listdir(origin_path)
for i in range(len(list_photo)):
img = os.path.join(origin_path, list_photo[i])
img = utils.read_cv2_img(img)
output_img, rec_img = sr_forward(img)
utils.save_image(
rec_img, os.path.join(img_path, 'resolution',
'resol_' + list_photo[i]))
'''
img = utils.read_cv2_img(img_path)
output_img, rec_img = sr_forward(img)
utils.save_image(output_img, 'output_face.jpg')
utils.save_image(rec_img, 'output_img.jpg')
'''