👤 Face recognition, and its limits · การรู้จำใบหน้า และขีดจำกัดของมัน · 人脸识别，以及它的极限

1 🧑‍🤝‍🧑 Two different jobs · สองงานที่ต่างกัน · 两件不同的事

2 📦 One install line · ติดตั้งบรรทัดเดียว · 一行安装命令

terminal

pip install "opencv-contrib-python==4.12.0.88"

3 🔍 Find a face · หาใบหน้า · 找到一张脸

find_faces.py

"""find_faces.py — is there a face in this picture, and where?

    python find_faces.py photo.jpg

It draws a box around every face it finds and saves the picture as found_photo.jpg.
No AI model is downloaded: the rules for what a face looks like ship with OpenCV.
"""
import sys
from pathlib import Path

import cv2

CASCADE = cv2.data.haarcascades + "haarcascade_frontalface_default.xml"


def find_faces(path):
    """Returns the picture, and a box (x, y, width, height) for every face."""
    picture = cv2.imread(str(path))
    if picture is None:
        raise SystemExit(f"cannot open {path}")
    grey = cv2.cvtColor(picture, cv2.COLOR_BGR2GRAY)      # the detector ignores colour
    detector = cv2.CascadeClassifier(CASCADE)
    faces = detector.detectMultiScale(grey,
                                      scaleFactor=1.1,    # look for faces 10% bigger each pass
                                      minNeighbors=5,     # higher = fewer false alarms
                                      minSize=(30, 30))   # ignore anything tiny
    return picture, faces


if __name__ == "__main__":
    path = Path(sys.argv[1])
    picture, faces = find_faces(path)
    print(f"{path.name}: {len(faces)} face(s)")
    for (x, y, w, h) in faces:
        print(f"   a face at x={x} y={y}, {w} by {h} pixels")
        cv2.rectangle(picture, (x, y), (x + w, y + h), (0, 255, 0), 2)
    out = path.with_name("found_" + path.name)
    cv2.imwrite(str(out), picture)
    print("saved", out.name)

four real photos, one run each

jay_chou_headshot.jpg: 1 face(s)
   a face at x=5 y=89, 167 by 167 pixels

michelle_yeoh_2015.jpg: 1 face(s)
   a face at x=138 y=273, 537 by 537 pixels

jay_chou_the_era_singapore_2010_concert.jpg: 0 face(s)

kungfudunkmoviepromo1.jpg: 8 face(s)
   a face at x=56 y=158, 141 by 141 pixels
   a face at x=757 y=240, 127 by 127 pixels
   a face at x=382 y=450, 142 by 142 pixels
   ... and five more

4 🙋 Give it photos, then a new photo · ให้รูป แล้วให้รูปใหม่ · 先给照片，再给一张新的

the folders

faces/
  jay_chou/       00.jpg 01.jpg ... 07.jpg
  michelle_yeoh/  00.jpg 01.jpg ... 07.jpg

who_is_this.py

"""who_is_this.py — learn a few faces, then name the person in a new photo.

    faces/
      ploy/    1.jpg 2.jpg 3.jpg ...     a few photos of each person
      nok/     1.jpg 2.jpg 3.jpg ...

    python who_is_this.py faces test.jpg

It prints a name and a distance. A LOW distance means a good match. Above the
threshold it says "I do not know", which is the answer you want for a stranger.
"""
import sys
from pathlib import Path

import cv2
import numpy as np

CASCADE = cv2.data.haarcascades + "haarcascade_frontalface_default.xml"
SIZE = (200, 200)          # every face is cut out and resized to this before learning
THRESHOLD = 56             # distance above this = not sure enough to give a name.
                           # Measured on 101 photos, not guessed. Read the page:
                           # this number is a compromise, not a solution.


def face_from(path):
    """The largest face in a photo, in grey, cut out and resized. None if there is none."""
    picture = cv2.imread(str(path), cv2.IMREAD_GRAYSCALE)
    if picture is None:
        return None
    faces = cv2.CascadeClassifier(CASCADE).detectMultiScale(picture, 1.1, 5, minSize=(30, 30))
    if len(faces) == 0:
        return None
    x, y, w, h = max(faces, key=lambda box: box[2] * box[3])
    return cv2.resize(picture[y:y + h, x:x + w], SIZE)


def learn(folder):
    """Train on faces/<name>/*.jpg. Returns the model and the list of names."""
    faces, labels, names = [], [], []
    for person in sorted(p for p in Path(folder).iterdir() if p.is_dir()):
        photos = sorted(list(person.glob("*.jpg")) + list(person.glob("*.png")))
        found = [face_from(photo) for photo in photos]
        found = [f for f in found if f is not None]
        if not found:
            print(f"  {person.name}: no face found in any photo, skipping")
            continue
        names.append(person.name)
        faces += found
        labels += [len(names) - 1] * len(found)
        print(f"  {person.name}: learned {len(found)} of {len(photos)} photos")
    model = cv2.face.LBPHFaceRecognizer_create()
    model.train(faces, np.array(labels))
    return model, names


def who(model, names, path):
    """Name the face in one photo, or say it does not know."""
    face = face_from(path)
    if face is None:
        return "no face found", None
    label, distance = model.predict(face)
    if distance > THRESHOLD:
        return "I do not know", distance
    return names[label], distance


if __name__ == "__main__":
    folder, test = sys.argv[1], sys.argv[2]
    print("learning:")
    model, names = learn(folder)
    print("knows:", ", ".join(names))
    name, distance = who(model, names, test)
    if distance is None:
        print(f"{Path(test).name}: {name}")
    else:
        print(f"{Path(test).name}: {name}  (distance {distance:.1f}, threshold {THRESHOLD})")

terminal

python who_is_this.py faces test_jay_chou.jpg

five photos it had never been trained on

test_jay_chou.jpg:        jay_chou       (distance 54.6, threshold 56)
test_michelle_yeoh.jpg:   michelle_yeoh  (distance 34.6, threshold 56)
hard_jay_chou.jpg:        I do not know  (distance 64.3, threshold 56)
stranger_zhang_ziyi.jpg:  I do not know  (distance 59.2, threshold 56)
stranger_jackie_chan.jpg: michelle_yeoh  (distance 54.0, threshold 56)

5 📏 How do you know it works? · จะรู้ได้อย่างไรว่ามันใช้ได้ · 你怎么知道它好用？

6 📷 The camera: find faces, name nobody · กล้อง: หาใบหน้า ไม่บอกชื่อ · 摄像头：只找脸，不认人

watch_camera.py

"""watch_camera.py - find faces in the laptop camera picture.

    python watch_camera.py

It draws a box around every face it can see, and counts them. It does not try to
say who anyone is. That is the other half of the job, and section 5 shows how
often it gets that wrong.

Press q to stop. Nothing is recorded: each frame is looked at and thrown away.
"""
import cv2

CAMERA = 0             # 0 is the built-in laptop camera. Try 1 for a plugged-in one.
MIN_FACE = 80          # ignore anything smaller than this, in pixels
CASCADE = cv2.data.haarcascades + "haarcascade_frontalface_default.xml"
detector = cv2.CascadeClassifier(CASCADE)


def find_faces(frame):
    """Every face in one camera frame, as a box (x, y, width, height)."""
    grey = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)      # the detector ignores colour
    return detector.detectMultiScale(grey,
                                     scaleFactor=1.1,
                                     minNeighbors=5,
                                     minSize=(MIN_FACE, MIN_FACE))


if __name__ == "__main__":
    camera = cv2.VideoCapture(CAMERA)
    if not camera.isOpened():
        raise SystemExit("cannot open the camera - is another program using it?")
    print("camera on. Press q to stop.")
    while True:
        ok, frame = camera.read()
        if not ok:
            break
        faces = find_faces(frame)
        for (x, y, w, h) in faces:
            cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 200, 0), 2)
            cv2.putText(frame, f"{w} by {h}", (x, y - 8),
                        cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 200, 0), 2)
        cv2.putText(frame, f"faces: {len(faces)}", (10, 30),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 200, 0), 2)
        cv2.imshow("find a face - press q to quit", frame)
        if cv2.waitKey(1) & 0xFF == ord("q"):
            break
    camera.release()
    cv2.destroyAllWindows()

terminal

python watch_camera.py

a real run on this laptop, with nobody in front of the camera

camera on. Press q to stop.
   faces: 0
(camera 640x480, 320 frames read, nothing recorded)

7 👉 Where it breaks, and what to try · มันพังตรงไหน และลองอะไรต่อ · 它在哪里出错，接下来试什么