Python-World · chavarera · Nov 16, 2021 · Sep 16, 2021 · Sep 16, 2021
diff --git a/projects/detect_align_faces/README.md b/projects/detect_align_faces/README.md
@@ -0,0 +1,25 @@
+# Detect and align faces
+
+This algorithm can detect the faces from picture and then align them.
+
+## Requirement
+
+**Dowload model parameters:**
+
+You should dowload the [model parameters](http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2) 
+from dlib, decompress and move to `./dat`
+
+**Installation:**
+
+```shell
+$ pip install -r requirements.txt
+```
+
+## Usage
+
+```shell
+$ python3 main.py [pic1, pic2...]
+```
+
+After the script finished, you will get some faces picture in the same
+directory.
diff --git a/projects/detect_align_faces/example.jpg b/projects/detect_align_faces/example.jpg
diff --git a/projects/detect_align_faces/main.py b/projects/detect_align_faces/main.py
@@ -0,0 +1,74 @@
+#!/usr/bin/env python3
+#
+# Copyright(C) 2021 wuyaoping
+#
+
+
+import numpy as np
+import os.path as osp
+import sys
+import cv2
+import dlib
+
+OUT_SIZE = (224, 224)
+LEFT_EYE_RANGE = (36, 42)
+RIGHT_EYE_RABGE = (42, 48)
+LEFT_EYE_POS = (0.35, 0.3815)
+DAT_PATH = "./dat/shape_predictor_68_face_landmarks.dat"
+
+
+def main(files):
+    detector = dlib.get_frontal_face_detector()
+    sp = dlib.shape_predictor(DAT_PATH)
+
+    for file in files:
+        img = cv2.imread(file, cv2.IMREAD_ANYCOLOR)
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+
+        faces = detect_align_faces(detector, sp, img)
+        for (idx, face) in enumerate(faces):
+            face = cv2.cvtColor(face, cv2.COLOR_RGB2BGR)
+            filename, ext = osp.splitext(file)
+            filename += '_face_{:03}'.format(idx) + ext
+            cv2.imwrite(filename, face)
+
+
+def detect_align_faces(detector, sp, img):
+    faces = detector(img, 1)
+    res = []
+    for face in faces:
+        shape = sp(img, face)
+        left, right = shape_to_pos(shape)
+        left_center = np.mean(left, axis=0)
+        right_center = np.mean(right, axis=0)
+
+        dx = right_center[0] - left_center[0]
+        dy = right_center[1] - left_center[1]
+        angle = np.degrees(np.arctan2(dy, dx))
+        dist = np.sqrt(dy ** 2 + dx ** 2)
+        out_dist = OUT_SIZE[0] * (1 - 2 * LEFT_EYE_POS[0])
+        scale = out_dist / dist
+        center = ((left_center + right_center) // 2).tolist()
+
+        mat = cv2.getRotationMatrix2D(center, angle, scale)
+        mat[0, 2] += (0.5 * OUT_SIZE[0] - center[0])
+        mat[1, 2] += (LEFT_EYE_POS[1] * OUT_SIZE[1] - center[1])
+        res_face = cv2.warpAffine(img, mat, OUT_SIZE, flags=cv2.INTER_CUBIC)
+        res.append(res_face)
+
+    return res
+
+
+def shape_to_pos(shape):
+    parts = []
+    for p in shape.parts():
+        parts.append((p.x, p.y))
+
+    left = parts[LEFT_EYE_RANGE[0]: LEFT_EYE_RANGE[-1]]
+    right = parts[RIGHT_EYE_RABGE[0]: RIGHT_EYE_RABGE[-1]]
+
+    return (np.array(left), np.array(right))
+
+
+if __name__ == '__main__':
+    main(sys.argv[1:])
diff --git a/projects/detect_align_faces/requirements.txt b/projects/detect_align_faces/requirements.txt
@@ -0,0 +1,3 @@
+dlib==19.22.1
+numpy==1.21.2
+opencv-python==4.5.3.56