Project — Smile Detector App

• Nest the smile cascade inside detected faces.
• React live: label, colour, or auto-snapshot on a smile.
• Tune the smile detector to reduce false triggers.
• Ship a complete, interactive vision app.

for each frame:
  detect faces (on the whole gray frame)
  for each face:
    crop the lower-half of the face   (smiles live near the mouth)
    run smile cascade on that region
    if a smile is found → react

Load both cascades

import cv2
face = cv2.CascadeClassifier(cv2.data.haarcascades +
                             "haarcascade_frontalface_default.xml")
smile = cv2.CascadeClassifier(cv2.data.haarcascades +
                              "haarcascade_smile.xml")

Smile detection needs strict params

The smile cascade is twitchy — it triggers on any mouth-shaped region. Use a high minNeighbors and a reasonable minSize to require a real smile:

smiles = smile.detectMultiScale(face_region_gray,
                                scaleFactor=1.7,
                                minNeighbors=22,    # strict!
                                minSize=(25, 25))

Search only the lower face

# within a detected face (x, y, w, h):
lower = gray[y + h // 2 : y + h, x : x + w]   # bottom half
smiles = smile.detectMultiScale(lower, 1.7, 22, minSize=(25, 25))
is_smiling = len(smiles) > 0

React

colour = (0, 255, 0) if is_smiling else (200, 200, 200)
label  = "smile :)" if is_smiling else "..."
cv2.rectangle(frame, (x, y), (x+w, y+h), colour, 2)
cv2.putText(frame, label, (x, y-8),
            cv2.FONT_HERSHEY_SIMPLEX, 0.7, colour, 2)

# smile_detector.py — full interactive app
import cv2, time

face = cv2.CascadeClassifier(cv2.data.haarcascades +
                             "haarcascade_frontalface_default.xml")
smile = cv2.CascadeClassifier(cv2.data.haarcascades +
                              "haarcascade_smile.xml")

cap = cv2.VideoCapture(0)
shots = 0
last_shot = 0

try:
    while True:
        ok, frame = cap.read()
        if not ok:
            break
        frame = cv2.flip(frame, 1)                 # selfie mirror
        gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)

        for (x, y, w, h) in face.detectMultiScale(gray, 1.2, 5, minSize=(80, 80)):
            lower = gray[y + h // 2 : y + h, x : x + w]
            smiles = smile.detectMultiScale(lower, 1.7, 22, minSize=(25, 25))
            smiling = len(smiles) > 0

            colour = (0, 255, 0) if smiling else (180, 180, 180)
            cv2.rectangle(frame, (x, y), (x + w, y + h), colour, 2)
            cv2.putText(frame, "smile :)" if smiling else "...",
                        (x, y - 8), cv2.FONT_HERSHEY_SIMPLEX, 0.7, colour, 2)

            # auto-snapshot on a smile, at most once per 2 seconds
            if smiling and time.time() - last_shot > 2:
                cv2.imwrite(f"smile_{shots}.jpg", frame)
                shots += 1
                last_shot = time.time()

        cv2.imshow("Smile Detector (q to quit)", frame)
        if cv2.waitKey(1) & 0xFF == ord("q"):
            break
finally:
    cap.release()
    cv2.destroyAllWindows()
    print(f"captured {shots} smiles")

Read the diff

Nested detection (smiles only in the lower face), strict smile params, a live colour/label reaction, and a debounced auto-snapshot (the 2-second cooldown stops it spamming 30 photos a second). This is a real, complete app built entirely from cascades you already know.

You shipped an interactive vision app by nesting detectors (faces → smiles in the lower face), using strict params to cut false triggers, reacting live, and debouncing the snapshot. No new APIs — just composition of the vision week. That closes computer vision; next we teach computers to read language.