trysem/vintager

1# Copyright (c) 2021 Justin Pinkney2 3import dlib4import numpy as np5import os6from PIL import Image7from PIL import ImageOps8from scipy.ndimage import gaussian_filter9import cv210 11 12MODEL_PATH = "shape_predictor_5_face_landmarks.dat"13detector = dlib.get_frontal_face_detector()14 15 16def align(image_in, face_index=0, output_size=256):17    try:18        image_in = ImageOps.exif_transpose(image_in)19    except:20        print("exif problem, not rotating")21 22    landmarks = list(get_landmarks(image_in))23    n_faces = len(landmarks)24    face_index = min(n_faces-1, face_index)25    if n_faces == 0:26        aligned_image = image_in27        quad = None28    else:29        aligned_image, quad = image_align(image_in, landmarks[face_index], output_size=output_size)30 31    return aligned_image, n_faces, quad32 33 34def composite_images(quad, img, output):35    """Composite an image into and output canvas according to transformed co-ords"""36    output = output.convert("RGBA")37    img = img.convert("RGBA")38    input_size = img.size39    src = np.array(((0, 0), (0, input_size[1]), input_size, (input_size[0], 0)), dtype=np.float32)40    dst = np.float32(quad)41    mtx = cv2.getPerspectiveTransform(dst, src)42    img = img.transform(output.size, Image.PERSPECTIVE, mtx.flatten(), Image.BILINEAR)43    output.alpha_composite(img)44 45    return output.convert("RGB")46    47 48def get_landmarks(image):49    """Get landmarks from PIL image"""50    shape_predictor = dlib.shape_predictor(MODEL_PATH)51 52    max_size = max(image.size)53    reduction_scale = int(max_size/512)54    if reduction_scale == 0:55        reduction_scale = 156    downscaled = image.reduce(reduction_scale)57    img = np.array(downscaled)58    detections = detector(img, 0)59    60    for detection in detections:61        try:62            face_landmarks = [(reduction_scale*item.x, reduction_scale*item.y) for item in shape_predictor(img, detection).parts()]63            yield face_landmarks64        except Exception as e:65            print(e)66 67 68def image_align(src_img, face_landmarks, output_size=512, transform_size=2048, enable_padding=True, x_scale=1, y_scale=1, em_scale=0.1, alpha=False):69        # Align function modified from ffhq-dataset70        # See https://github.com/NVlabs/ffhq-dataset for license71 72        lm = np.array(face_landmarks)73        lm_eye_left      = lm[2:3]  # left-clockwise74        lm_eye_right     = lm[0:1]  # left-clockwise75 76        # Calculate auxiliary vectors.77        eye_left     = np.mean(lm_eye_left, axis=0)78        eye_right    = np.mean(lm_eye_right, axis=0)79        eye_avg      = (eye_left + eye_right) * 0.580        eye_to_eye   = 0.71*(eye_right - eye_left)81        mouth_avg    = lm[4]82        eye_to_mouth = 1.35*(mouth_avg - eye_avg)83 84        # Choose oriented crop rectangle.85        x = eye_to_eye.copy()86        x /= np.hypot(*x)87        x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8)88        x *= x_scale89        y = np.flipud(x) * [-y_scale, y_scale]90        c = eye_avg + eye_to_mouth * em_scale91        quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])92        quad_orig = quad.copy()93        qsize = np.hypot(*x) * 2        94        95        img = src_img.convert('RGBA').convert('RGB')96 97        # Shrink.98        shrink = int(np.floor(qsize / output_size * 0.5))99        if shrink > 1:100            rsize = (int(np.rint(float(img.size[0]) / shrink)), int(np.rint(float(img.size[1]) / shrink)))101            img = img.resize(rsize, Image.ANTIALIAS)102            quad /= shrink103            qsize /= shrink104 105        # Crop.106        border = max(int(np.rint(qsize * 0.1)), 3)107        crop = (int(np.floor(min(quad[:,0]))), int(np.floor(min(quad[:,1]))), int(np.ceil(max(quad[:,0]))), int(np.ceil(max(quad[:,1]))))108        crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, img.size[0]), min(crop[3] + border, img.size[1]))109        if crop[2] - crop[0] < img.size[0] or crop[3] - crop[1] < img.size[1]:110            img = img.crop(crop)111            quad -= crop[0:2]112 113        # Pad.114        pad = (int(np.floor(min(quad[:,0]))), int(np.floor(min(quad[:,1]))), int(np.ceil(max(quad[:,0]))), int(np.ceil(max(quad[:,1]))))115        pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - img.size[0] + border, 0), max(pad[3] - img.size[1] + border, 0))116        if enable_padding and max(pad) > border - 4:117            pad = np.maximum(pad, int(np.rint(qsize * 0.3)))118            img = np.pad(np.float32(img), ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')119            h, w, _ = img.shape120            y, x, _ = np.ogrid[:h, :w, :1]121            mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0], np.float32(w-1-x) / pad[2]), 1.0 - np.minimum(np.float32(y) / pad[1], np.float32(h-1-y) / pad[3]))122            blur = qsize * 0.02123            img += (gaussian_filter(img, [blur, blur, 0]) - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)124            img += (np.median(img, axis=(0,1)) - img) * np.clip(mask, 0.0, 1.0)125            img = np.uint8(np.clip(np.rint(img), 0, 255))126            if alpha:127                mask = 1-np.clip(3.0 * mask, 0.0, 1.0)128                mask = np.uint8(np.clip(np.rint(mask*255), 0, 255))129                img = np.concatenate((img, mask), axis=2)130                img = Image.fromarray(img, 'RGBA')131            else:132                img = Image.fromarray(img, 'RGB')133            quad += pad[:2]134 135        # Transform.136        img = img.transform((transform_size, transform_size), Image.QUAD, (quad + 0.5).flatten(), Image.BILINEAR)137        if output_size < transform_size:138            img = img.resize((output_size, output_size), Image.ANTIALIAS)139 140        return img, quad_orig141