src/face_proportions.py

import dlib
import cv2
import os
import numpy as np
import imutils
import urllib.request
import bz2
from src.cv_utils import get_image, resize_image_height
from typing import List, Union
from PIL import Image as PILImage


def download_shape_predictor_model() -> str:
    model_url = "http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2"
    local_bz2_path = "models/shape_predictor_68_face_landmarks.dat.bz2"
    local_dat_path = "models/shape_predictor_68_face_landmarks.dat"

    if not os.path.exists(local_dat_path):
        os.makedirs("models", exist_ok=True)
        print("Downloading shape predictor model...")
        urllib.request.urlretrieve(model_url, local_bz2_path)

        with bz2.open(local_bz2_path, "rb") as f_in, open(local_dat_path, "wb") as f_out:
            f_out.write(f_in.read())
        print("Model extracted.")

    return local_dat_path


class GetFaceProportions:
    def __init__(self):
        pass

    @staticmethod
    def preprocess_image(image: np.array) -> np.array:
        image = imutils.resize(image, width=500)
        gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
        return gray_image

    @staticmethod
    def detect_face_landmarks(gray_image: np.array) -> List[Union[np.array, np.array]]:
        detector = dlib.get_frontal_face_detector()
        predictor_path = download_shape_predictor_model()
        predictor = dlib.shape_predictor(predictor_path)
        rects = detector(gray_image, 1)
        for rect in rects:
            shape = predictor(gray_image, rect)
            shape = np.array(
                [(shape.part(i).x, shape.part(i).y) for i in range(shape.num_parts)]
            )

            # Draw facial landmarks
            for (x, y) in shape:
                cv2.circle(gray_image, (x, y), 2, (0, 255, 0), -1)

        return shape, gray_image

    @staticmethod
    def compute_golden_ratios(shape: np.array) -> dict:
        top_mouth, middle_mouth, bottom_mouth = shape[51], shape[62], shape[57]
        top_nose, bottom_nose = shape[27], shape[33]
        bottom_chin = shape[8]

        # 1
        top_nose_to_middle_mouth_dist = np.linalg.norm(
            top_nose - middle_mouth
        )  # euclidean distance
        middle_mouth_to_bottom_chin_dist = np.linalg.norm(middle_mouth - bottom_chin)
        ratio_top_nose_to_middle_mouth_vs_middle_mouth_to_bottom_chin = (
            top_nose_to_middle_mouth_dist / middle_mouth_to_bottom_chin_dist
        )

        # 2
        top_mouth_to_middle_mouth_dist = np.linalg.norm(top_mouth - middle_mouth)
        middle_mouth_to_bottom_mouth_dist = np.linalg.norm(middle_mouth - bottom_mouth)
        ratio_middle_mouth_to_bottom_mouth_vs_top_mouth_to_middle_mouth = (
            middle_mouth_to_bottom_mouth_dist / top_mouth_to_middle_mouth_dist
        )

        golden_ratios = {
            "top_of_nose_to_middle_of_mouth_vs_middle_mouth_to_bottom_of_chin": ratio_top_nose_to_middle_mouth_vs_middle_mouth_to_bottom_chin,
            "middle_of_mouth_to_bottom_of_mouth_vs_top_of_mouth_to_middle_of_mouth": ratio_middle_mouth_to_bottom_mouth_vs_top_mouth_to_middle_mouth,
        }
        return golden_ratios

    @staticmethod
    def compute_equal_ratios(shape: np.array) -> dict:
        (
            left_side_left_eye,
            right_side_left_eye,
            left_side_right_eye,
            right_side_right_eye,
        ) = (shape[36], shape[39], shape[42], shape[45])
        left_eye_top, left_eye_bottom, right_eye_top, right_eye_bottom = (
            shape[37],
            shape[41],
            shape[44],
            shape[46],
        )
        left_eyebrow_top, right_eyebrow_top = shape[19], shape[24]
        left_eye_center = np.mean([shape[37], shape[38], shape[41], shape[40]], axis=0)
        right_eye_center = np.mean([shape[43], shape[44], shape[47], shape[46]], axis=0)
        left_mouth, right_mouth = shape[48], shape[54]

        # 1
        left_eye_dist = np.linalg.norm(left_side_left_eye - right_side_left_eye)
        right_eye_dist = np.linalg.norm(left_side_right_eye - right_side_right_eye)
        average_eye_dist = (left_eye_dist + right_eye_dist) / 2
        between_eye_dist = np.linalg.norm(right_side_left_eye - left_side_right_eye)
        ratio_eyes_width_vs_between_eye = average_eye_dist / between_eye_dist

        # 2
        left_eye_to_eyebrow_dist = np.linalg.norm(left_eyebrow_top - left_eye_top)
        right_eye_to_eyebrow_dist = np.linalg.norm(right_eyebrow_top - right_eye_top)
        eye_to_eyebrow_dist = (left_eye_to_eyebrow_dist + right_eye_to_eyebrow_dist) / 2
        left_eye_height = np.linalg.norm(left_eye_top - left_eye_bottom)
        right_eye_height = np.linalg.norm(right_eye_top - right_eye_bottom)
        eye_height = (left_eye_height + right_eye_height) / 2
        ratio_eye_to_eyebrow_vs_eye_height = eye_to_eyebrow_dist / eye_height

        # 3
        left_to_right_eye_center_dist = np.linalg.norm(
            left_eye_center - right_eye_center
        )
        mouth_width = np.linalg.norm(left_mouth - right_mouth)
        ratio_left_to_right_eye_center_vs_mouth_width = (
            left_to_right_eye_center_dist / mouth_width
        )

        equal_ratios = {
            "eye_width_vs_distance_between_eyes": ratio_eyes_width_vs_between_eye,
            "eye_to_eyebrows_vs_eye_height": ratio_eye_to_eyebrow_vs_eye_height,
            "center_of_left_to_right_eye_vs_mouth_width": ratio_left_to_right_eye_center_vs_mouth_width,
        }
        return equal_ratios

    def main(self, image_input):
        image = get_image(image_input)
        gray_image = self.preprocess_image(image)
        shape, image = self.detect_face_landmarks(gray_image)
        golden_ratios = self.compute_golden_ratios(shape)
        golden_ratios = {k: round(v, 2) for k, v in golden_ratios.items()}
        equal_ratios = self.compute_equal_ratios(shape)
        equal_ratios = {k: round(v, 2) for k, v in equal_ratios.items()}
        image = PILImage.fromarray(image)
        image = resize_image_height(image, new_height=300)
        ratios = {**golden_ratios, **equal_ratios}
        return ratios, image


if __name__ == "__main__":
    path_to_images = "data/"
    image_files = os.listdir(path_to_images)
    for image in image_files:
        print(image)
        results = GetFaceProportions().main(path_to_images + image)
        print(results)