前言
看完文章您将学会：
paddlehub的使用方法
如何用cv2加载图片并保存
如何用cv2逐帧加载视频以及将图片逐帧保存成视频
如何将png格式的图片放入另一张图片
本文涉及paddlehub的人脸检测、图像分割和图像生成三个部分
详细的文档请访问： https://www.paddlepaddle.org.cn/hub
简介
本项目通过人脸检测将人脸遮挡实现打马赛的功能，同时通过风格转换和抠图将人物放置在新的背景下实现换背景的效果。
处理后的图片效果图：
paddlehub安装
pip install paddlehub --upgrade -i https://mirror.baidu.com/pypi/simple
paddlehub 中的模型对版本有要求
请安装最新版本的paddlehub
或者根据需要指定安装版本：
hub install name==version
功能实现
引入库
代码如下：
import paddle
import paddlehub as hub
import numpy as np
from PIL import Image, ImageFilter, ImageDraw
import cv2, os
import matplotlib.pyplot as plt
%matplotlib inline
print(paddle.__version__)
用到的hub库
代码如下：
#用于人脸检测
face_detection = hub.Module(name="ultra_light_fast_generic_face_detector_1mb_640")
#用于风格转换
stylepro_artistic = hub.Module(name="stylepro_artistic")
#用于扣出人物
humanseg = hub.Module(name="deeplabv3p_xception65_humanseg")
可以查看一下数据的输出格式，本文中默认只考虑一个物体的情况：
# 查看一下使用方法及输出格式， 这里默认一张图片中只有一个目标物
face_detection_res = face_detection.face_detection(images=[cv2.imread('./work/me1.jpg')],
paths=None,
batch_size=1,
use_gpu=False,
visualization=False,
output_dir=None,
confs_threshold=0.5)
# [0]['data']
result = stylepro_artistic.style_transfer( images=[{'content': cv2.imread('./dog.png'),
'styles': [cv2.imread('./style/style1.jpg'), cv2.imread('./style/style2.jpg'),cv2.imread('./style/style3.jpg'),cv2.imread('./style/style4.jpg'),cv2.imread('./style/style5.jpg')]
}],
visualization=True)
# [0]['data']
seg_res = humanseg.segment(images=[cv2.imread('./work/me1.jpg')],
paths=None,
batch_size=1,
use_gpu=False,
visualization=True,
output_dir='humanseg_output')
# [0]['data']
每帧的图像处理
通过修改precess_img可以实现不同的处理效果。
def process_img(frame_bgr, index=0):
'''
输入一张图片shape=[H, W, C] 通道为bgr格式
'''
ratio = 1.4
y_offset = 30
num_fps = 150
t = 1
h, w = frame_bgr.shape[:2]
size = frame_bgr.shape[:2]
face_detector = MyFaceDetector()
face_detection_res = face_detector.face_detection(images=[frame_bgr], use_gpu=False, visualization=False)
for box_dict in face_detection_res[0]['data']:
box_xyxy = ( int(box_dict['left']), int(box_dict['top']),
int(box_dict['right']), int(box_dict['bottom']) )
box_xywh = int((box_xyxy[0]+box_xyxy[2])/2), int((box_xyxy[1]+box_xyxy[3])/2), box_xyxy[2]-box_xyxy[0], box_xyxy[3]-box_xyxy[1]
correct_box_xywh = box_xywh[0], box_xywh[1]-y_offset, int(box_xywh[2]*ratio), int(box_xywh[3]*ratio)
#真实框xyxy
box = int(correct_box_xywh[0]-correct_box_xywh[2]/2) if int(correct_box_xywh[0]-correct_box_xywh[2]/2) >= 0 else 0, \
int(correct_box_xywh[1]-correct_box_xywh[3]/2) if int(correct_box_xywh[1]-correct_box_xywh[3]/2) >= 0 else 0, \
int(correct_box_xywh[0]+correct_box_xywh[2]/2) if int(correct_box_xywh[0]+correct_box_xywh[2]/2) <= size[1] else size[1], \
int(correct_box_xywh[1]+correct_box_xywh[3]/2) if int(correct_box_xywh[1]+correct_box_xywh[3]/2) <= size[0] else size[0]
dog = cv2.imread('./dog.png', -1) # -1 读取alpha通道
dog = cv2.resize(dog, ( box[2]-box[0] if (box[2]-box[0])>0 else 1 , box[3]-box[1] if (box[3]-box[1])>0 else 1) )
dog_alpha = dog[:,:,3] != 0
dog_alpha = np.repeat(dog_alpha[:,:,np.newaxis], axis=2, repeats=3)
human_alpha = humanseg.segmentation(images=[frame_bgr],
paths=None,
batch_size=1,
use_gpu=False,
visualization=False,
output_dir='humanseg_output')[0]['data']
human_alpha = np.repeat(human_alpha[:,:,np.newaxis], axis=2, repeats=3) != 0
if index <= num_fps:
bg = cv2.imread('./bg1.png')
bg = cv2.resize(bg, (w, h))
elif index > num_fps and index < (num_fps + t*30):
beta = (index-num_fps)/30
bg = beta * cv2.imread('./bg2.jpg')/255. + (1 - beta) * cv2.imread('./bg1.jpg')/255.
bg = bg * 255
bg = bg.astype('uint8')
bg = cv2.resize(bg, (w, h))
else:
bg = cv2.imread('./bg2.png')
bg = cv2.resize(bg, (w, h))
# 加了这两项后运算时间会大大延长
# frame_bgr = stylepro_artistic.style_transfer(images=[{'content': frame_bgr,
# 'styles': [cv2.imread('./style/style1.jpg')]
# }], use_gpu=False,
# visualization=False)[0]['data']
# dog = stylepro_artistic.style_transfer(images=[{'content': dog[:,:,:3],
# 'styles': [cv2.imread('./style/style1.jpg')]
# }], use_gpu=False,
# visualization=False)[0]['data']
bg[human_alpha] = frame_bgr[human_alpha] #根据alpah矩阵赋值
bg[box[1]:box[1]+dog.shape[0], box[0]:box[0]+dog.shape[1], :][dog_alpha] = dog[:,:,:3][dog_alpha]
# dog = cv2.imread('./dog.png', -1)
# point_boxwh = (point[0], point[1] - (box[3]-box[1])//2 , box[2]-box[0], box[3]-box[1])
#
# point_box = point_boxwh[0]-point_boxwh[2]//2 if (point_boxwh[0]-point_boxwh[2]//2) >= 0 else 0, \
# point_boxwh[1]-point_boxwh[3]//2 if (point_boxwh[1]-point_boxwh[3]//2) >= 0 else 0, \
# point_boxwh[0]+point_boxwh[2]//2 if (point_boxwh[0]+point_boxwh[2]//2) <= size[1] else size[1], \
# point_boxwh[1]+point_boxwh[3]//2 if (point_boxwh[1]+point_boxwh[3]//2) <= size[0] else size[0]
# dog = cv2.resize(dog, ( point_box[2]-point_box[0] if (point_box[2]-point_box[0])>0 else 1, point_box[3]-point_box[1] if (point_box[3]-point_box[1])>0 else 1) )
# alpha_channel = dog[:,:,3] != 0
# alpha_channel = np.repeat(alpha_channel[:,:,np.newaxis], axis=2, repeats=3)
# # assert point_box[1]:point_box[1]+dog.shape[0], point_box[0]: point_box[0]+dog.shape[1]
# frame_bgr[point_box[1]:point_box[1]+dog.shape[0], point_box[0]: point_box[0]+dog.shape[1],:][alpha_channel] = dog[:,:,:3][alpha_channel]
return bg
def CutVideo2Image(video_path, img_path):
#将视频输出为图像
#video_path为输入视频文件路径
#img_path为输出图像文件夹路径
cap = cv2.VideoCapture(video_path)
index = 0
while(True):
ret,frame = cap.read()
if ret:
cv2.imwrite(img_path+'/%d.jpg'%index, frame)
index += 1
else:
break
cap.release()
class MyFaceDetector(object):
"""
自定义人脸检测器
基于PaddleHub人脸检测模型ultra_light_fast_generic_face_detector_1mb_640，加强稳定人脸检测框
"""
def __init__(self):
self.module = hub.Module(name="ultra_light_fast_generic_face_detector_1mb_640")
self.alpha = 0.75
self.start_flag =1
def face_detection(self,images, use_gpu=False, visualization=False):
# 使用GPU运行，use_gpu=True，并且在运行整个教程代码之前设置CUDA_VISIBLE_DEVICES环境变量
result = self.module.face_detection(images=images, use_gpu=use_gpu, visualization=visualization)
if not result[0]['data']:
return result
face = result[0]['data'][0]
if self.start_flag == 1:
self.left_s = result[0]['data'][0]['left']
self.right_s = result[0]['data'][0]['right']
self.top_s = result[0]['data'][0]['top']
self.bottom_s = result[0]['data'][0]['bottom']
self.start_flag=0
else:
# 加权平均上一帧和当前帧人脸检测框位置，以稳定人脸检测框
self.left_s = self.alpha * self.left_s + (1-self.alpha) * face['left']
self.right_s = self.alpha * self.right_s + (1-self.alpha) * face['right']
self.top_s = self.alpha * self.top_s + (1-self.alpha) * face['top']
self.bottom_s = self.alpha * self.bottom_s + (1-self.alpha) * face['bottom']
result[0]['data'][0]['left'] = self.left_s
result[0]['data'][0]['right'] = self.right_s
result[0]['data'][0]['top'] = self.top_s
result[0]['data'][0]['bottom'] = self.bottom_s
return result
结果输出
def generate_image():
# 打开摄像头
# capture = cv2.VideoCapture(0)
capture = cv2.VideoCapture('./test.mp4')
fps = capture.get(cv2.CAP_PROP_FPS)
size = (int(capture.get(cv2.CAP_PROP_FRAME_WIDTH)),
int(capture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
# 将预测结果写成视频
video_writer = cv2.VideoWriter('result.mp4', cv2.VideoWriter_fourcc(*'mp4v'), fps, size)
print(f'fps={fps}, size={size}')
index = 0
while True:
# frame_rgb即视频的一帧数据
ret, frame_bgr = capture.read() #从capture中读取帧
# 按q键即可退出
# cv2.imwrite('./work/me1.jpg', frame_bgr)
# break
if cv2.waitKey(1) & 0xFF == ord('q'):
break
if frame_bgr is None:
break
# cv2.imwrite('./work'+'/%d.jpg'%index, frame_bgr)
index += 1
#图像处理
frame_bgr = process_img(frame_bgr, index)
video_writer.write(frame_bgr) #写入帧
# frame_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_RGB2BGR) # cv2.COLOR_RGB2BGR就是把0, 2 通道互换
# yield frame_rgb
capture.release()
video_writer.release()
cv2.destroyAllWindows()
generate_image()
总结
视频中人物的抠图效果需要提升，可以通过在cv2中进一步处理提升画面效果。
另外声音需要后期合成后单独加入，后续我会想办法把声音加上。
总的来看，paddlehub将一些主流的模型集成起来做成相应的接口，当需要时直接调用即可，使用起来也非常方便。这极大的降低了网络的使用门槛，只需要少量的代码即可实现复杂的功能。
914
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
