一起养成写作习惯!这是我参与「掘金日新计划 · 4 月更文挑战」的第26天,点击查看活动详情。
1.关键点检测
人脸关键点检测是人脸识别和分析领域中的关键一步,它是诸如自动人脸识别、表情分析、三维人脸重建及三维动画等其它人脸相关问题的前提和突破口。该 PaddleHub Module 的模型转换自 github.com/lsy17096535… ,支持同一张图中的多个人脸检测。
2.检测关键点
很简单,通过PaddleHub的face_landmark_localization模块进行关键点检测
import paddlehub as hub
import cv2
import os
face_landmark = hub.Module(name="face_landmark_localization")
result = face_landmark.keypoint_detection(images=[cv2.imread('1.jpg')], output_dir='face_landmark_output',
visualization=False)
# 左上:2
# 右上:16
# 最下中:9
# print(result)
data = result[0]['data'][0]
print(data)
print(len(data))
print("左上:2 ", data[1])
print("右上:16 ", data[15])
print("最下中:9 ", data[8])
width = int(data[15][0] - data[1][0])+50
height = int(data[8][1] - data[29][1])
# dx = int(data[1][0] + width / 2)
# dy = int(data[8][1] - height / 2)
dx = int(data[1][0])-10
dy = int(data[1][1])
print(width, height, dx, dy)
复制代码
3.合成
主要是根据检测的脸的大小resize口罩,最后进行图层合并
# 第二步:合成
def merge(face_pic, w, h, dx, dy):
mask_path = 'mask.png'
face = cv2.imread(face_pic, cv2.IMREAD_COLOR) # 捕获图像1
print('face pic shape: ', face.shape)
# 口罩
# IMREAD_UNCHANGED If set, return the loaded image as is (with alpha channel, otherwise it gets cropped).
# 因此Png必须是4通道的IMREAD_UNCHANGED
mask = cv2.imread(mask_path, cv2.IMREAD_UNCHANGED)
# mask = cv2.resize(mask, [w, h])
mask = cv2.resize(mask, (w, h))
rows, cols, channels = mask.shape
print('mask shape: ', rows, cols, channels)
roi = face[dy:dy + rows, dx:dx + cols]
print('dx + rows, dy + cols', dy + rows, dx + cols)
for i in range(rows):
for j in range(cols):
if not (mask[i, j][3] == 0): # 透明的意思
roi[i, j][0] = mask[i, j][0]
roi[i, j][1] = mask[i, j][1]
roi[i, j][2] = mask[i, j][2]
face[dy:dy + rows, dx:dx + cols] = roi
merge_img_path = 're_' + face_pic
cv2.imwrite(merge_img_path, face)
return 're_' + face_pic
merge('1.jpg', width, height, dx, dy)
复制代码