ちょっと前までGANにハマっていて、その時に作った遺産です。
kaggleのページに公開しました。
もし著作権的にアウトであれば、削除します。

内容はSafebooruからスクレイピングし、顔部分をトリミングした約500,000枚の画像です。
画像のサイズは $256 \times 256$ 以上で、揃えていません（できるだけ高解像度にしてあります）。

他のデータセットと違う特徴として、目の位置を検出することで、顔の位置や角度を揃えてあります。
このことでGANなどで学習を行うとき、安定しやすくなっているのではないかと思います。
また、誤検出などで生じた、顔以外の画像、位置のずれた画像などはSVMで学習させて弾いています。

顔の検出は有名なlbpcascade_animeface.xmlを使用しました。
目の検出には自作の検出器を使っています（昔公開したOpenCVの物体検出の訓練データを作るためのツールを使いました）。
Safebooruからのスクレイピングには参考にしたページがあった気がするのですが、忘れてしまいました。

コード

スクレイピング

scraping.py

import urllib.request
import xml.etree.ElementTree
import cv2
import os
import numpy
import math
import PIL.Image
import io
import concurrent.futures
import time

minimum_size = 256
target_eye_distance = 0.15

output_directory = 'D:/safebooru_face/'


def reject_image(image):
    image = cv2.resize(image, (256, 256), interpolation=cv2.INTER_AREA)
    reshaped_image = image.reshape(-1, 3)
    cov_image = numpy.cov(reshaped_image, rowvar=False)
    eigen_values = numpy.linalg.eigvals(cov_image)
    eigen_values = numpy.sort(eigen_values)

    is_monotone = eigen_values[2] > eigen_values[1] * 1000
    return is_monotone


def trim_image(image, face_classifier, eye_classifier):
    result_images = []

    gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    gray_image = cv2.equalizeHist(gray_image)

    face_rects = face_classifier.detectMultiScale(gray_image,
                                                  scaleFactor=1.02,
                                                  minNeighbors=2,
                                                  minSize=(minimum_size, minimum_size))

    for (x, y, w, h) in face_rects:
        center_x = x + w / 2
        center_y = y + h / 2
        trim_size = max(w, h)

        x_left = int(center_x - trim_size / 2)
        x_right = int(center_x + trim_size / 2)
        y_top = int(center_y - trim_size / 2)
        y_bottom = int(center_y + trim_size / 2)

        gray_face_roi = gray_image[y_top:y_bottom, x_left:x_right]

        for min_neighbors in [10, 8, 6, 5, 4, 3, 2, 1, 0]:
            eye_rects = eye_classifier.detectMultiScale(gray_face_roi,
                                                        scaleFactor=1.02,
                                                        minNeighbors=min_neighbors,
                                                        minSize=(trim_size // 32, trim_size // 32))

            if len(eye_rects) == 2:
                break
        else:
            continue

        eye_positions = []
        for (x, y, w, h) in eye_rects:
            eye_x = x_left + x + w / 2
            eye_y = y_top + y + h / 2

            if x_left < eye_x and eye_x < x_right and y_top < eye_y and eye_y < y_bottom:
                eye_positions.append(numpy.array([eye_x, eye_y]))

        left_eye = eye_positions[0] if eye_positions[0][0] < eye_positions[1][0] else eye_positions[1]
        right_eye = eye_positions[1] if eye_positions[0][0] < eye_positions[1][0] else eye_positions[0]

        eye_center = (left_eye + right_eye) / 2
        face_angle = math.degrees(math.atan2(right_eye[1] - left_eye[1], right_eye[0] - left_eye[0]))
        eye_distance = math.sqrt(numpy.sum((left_eye - right_eye) ** 2))

        if eye_distance < 5:
            continue

        scale = (target_eye_distance * trim_size) / eye_distance
        trim_size /= scale

        if trim_size < minimum_size:
            continue

        rotation_matrix = cv2.getRotationMatrix2D((eye_center[0], eye_center[1]), face_angle, 1.0)
        rotation_matrix[0, 2] += trim_size / 2 - eye_center[0]
        rotation_matrix[1, 2] += trim_size / 2 - eye_center[1]

        trimmed_image = cv2.warpAffine(image, rotation_matrix, (int(trim_size), int(trim_size)), flags=cv2.INTER_LANCZOS4, borderValue=(255, 255, 255))

        result_images.append(trimmed_image)

    return result_images


def download_image(file_url):
    try:
        bin_image = io.BytesIO(urllib.request.urlopen(file_url).read())
        pil_image = PIL.Image.open(bin_image)

        image = numpy.array(pil_image, dtype=numpy.uint8)

        if image.shape[2] == 2:
            image = cv2.cvtColor(image, cv2.COLOR_GRAY2BGR)
        elif image.shape[2] == 3:
            image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
        elif image.shape[2] == 4:
            image = cv2.cvtColor(image, cv2.COLOR_RGBA2BGR)
        else:
            return

        return image, file_url
    except Exception as e:
        print(e)


if __name__ == '__main__':
    os.makedirs(output_directory, exist_ok=True)

    tags = [
        '-grayscale', '-greyscale', '-no_humans', '-glasses', '-furry'
    ]

    face_classifier = cv2.CascadeClassifier('lbpcascade_animeface.xml')
    eye_classifier = cv2.CascadeClassifier('lbpcascade_animeeye4.xml')

    for page_number in range(100000):
        print('Page : {}'.format(page_number))

        url = 'https://safebooru.org/index.php?page=dapi&s=post&q=index&pid={pid}&tags={tags}'.format(
            pid=page_number,
            tags='+'.join(tags)
        )

        while True:
            try:
                # print('Request.')
                with urllib.request.urlopen(url) as response:
                    content = response.read()
                break
            except Exception as e:
                print('Failed.')
                print(e)
                time.sleep(1.0)
                continue

        print('Success!')

        posts_xml = xml.etree.ElementTree.fromstring(content)

        print('n_posts : {}'.format(len([post for post in posts_xml.iter('post')])))

        file_urls = []

        for post in posts_xml.iter('post'):
            if int(post.get('width')) < minimum_size or int(post.get('height')) < minimum_size:
                continue

            if int(post.get('width')) * int(post.get('height')) > 3000 * 3000:
                continue

            file_url = post.get('file_url')
            extention = os.path.splitext(os.path.basename(file_url))[1]

            if extention == '.gif':
                continue

            file_urls.append(file_url)

        with concurrent.futures.ThreadPoolExecutor() as executor:
            futures = [executor.submit(download_image, file_url) for file_url in file_urls]

            for future in futures:
                try:
                    image, file_url = future.result()
                    file_name = os.path.splitext(os.path.basename(file_url))[0]

                    if reject_image(image):
                        continue

                    trimmed_images = trim_image(image, face_classifier, eye_classifier)

                    for image_counter, trimmed_image in enumerate(trimmed_images):
                        output_path = os.path.join(output_directory, file_name + '_' + str(image_counter) + '.png')
                        cv2.imwrite(output_path, trimmed_image)

                        print('{} -> {}'.format(file_url, output_path))

                except Exception as e:
                    print(e)

SVMによる誤検出の検出

svm.py

import cv2
import glob
import numpy
import sklearn.ensemble
import sklearn.svm
import skimage.feature
import sklearn.pipeline
import sklearn.decomposition
import sklearn.preprocessing
import os
import shutil
import pickle
import traceback


def split_list(input_list, n):
    return [input_list[i:i + n] for i in range(0, len(input_list), n)]


def preprocess(file_name):
    image = cv2.imread(file_name)

    if image is None:
        raise OSError('File is not found.')

    image = image.astype(numpy.float)
    image_size = min(image.shape[0], image.shape[1])

    left = image.shape[0] // 2 - image_size // 4
    right = image.shape[0] // 2 + image_size // 4
    top = image.shape[1] // 2 - image_size // 4
    bottom = image.shape[1] // 2 + image_size // 4
    image = image[left:right, top:bottom, :]

    image = cv2.resize(image, (256, 256), interpolation=cv2.INTER_AREA)

    hog = skimage.feature.hog(image, pixels_per_cell=(32, 32)).ravel()
    small_image = cv2.resize(image, (8, 8), interpolation=cv2.INTER_AREA).ravel()
    return numpy.concatenate([hog, small_image])


if __name__ == '__main__':
    true_dir = 'D:/safebooru_class/ok_candidate'
    false_dir = 'D:/safebooru_class/ng_candidate'
    input_dir = 'D:/safebooru_face'
    ok_dir = 'D:/safebooru_class/ok'
    ng_dir = 'D:/safebooru_class/ng'
    ambiguous_dir = 'D:/safebooru_class/ambiguous'
    dataset_file = 'D:/safebooru_class/dataset'

    ok_files = glob.glob(ok_dir + '/*')  # + glob.glob(true_dir + '/*')
    ng_files = glob.glob(ng_dir + '/*')  # + glob.glob(false_dir + '/*')
    # input_files = glob.glob(input_dir + '/**', recursive=True)

    all_files = ok_files + ng_files

    if os.path.exists(dataset_file):
        with open(dataset_file, 'rb') as f:
            dataset = pickle.load(f)
    else:
        dataset = {}

    dataset_changed = False

    for i, file in enumerate(all_files):
        if os.path.basename(file) not in dataset.keys():
            print(f'Process : {file} ({i} / {len(all_files)})')

            try:
                dataset[os.path.basename(file)] = preprocess(file)
                dataset_changed = True
            except OSError:
                continue

    if dataset_changed:
        print('Writing Dataset...')
        with open(dataset_file, 'wb') as f:
            pickle.dump(dataset, f)
        print('Done.')

    for iter_counter in range(1):
        ok_files = glob.glob(ok_dir + '/*')  # + glob.glob(true_dir + '/*')
        ng_files = glob.glob(ng_dir + '/*')  # + glob.glob(false_dir + '/*')
        input_files = glob.glob(input_dir + '/**', recursive=True)

        if len(input_files) < 100:
            break

        x = []
        y = []

        x.extend([dataset[os.path.basename(file)] for file in ok_files])
        y.extend([1] * len(ok_files))

        print('End OK Preprocess.')

        x.extend([dataset[os.path.basename(file)] for file in ng_files])
        y.extend([0] * len(ng_files))

        print('End NG Preprocess.')

        x = numpy.stack(x, axis=0)
        y = numpy.stack(y, axis=0)

        # classifier = sklearn.ensemble.VotingClassifier([
        #     ('randomforest', sklearn.ensemble.RandomForestClassifier()),
        #     ('svm', sklearn.svm.SVC(probability=True)),
        #     # ('nusvm', sklearn.svm.NuSVC(probability=True)),
        #     ('gradientboost', sklearn.ensemble.GradientBoostingClassifier()),
        #     ('adaboost', sklearn.ensemble.AdaBoostClassifier())
        # ], voting='soft')

        # classifier = sklearn.tree.DecisionTreeClassifier()
        # classifier = sklearn.svm.SVC(probability=True)

        classifier = sklearn.pipeline.Pipeline(steps=[
            ('scale1', sklearn.preprocessing.RobustScaler()),
            ('pca', sklearn.decomposition.PCA(n_components=100, whiten=True)),
            ('svc', sklearn.svm.SVC(probability=True)),
        ])

        classifier.fit(x, y)

        print('End Train.')

        for split_input_files in split_list(input_files, 1000):
            u = []
            new_input_files = []
            for i, file in enumerate(split_input_files):
                if not os.path.exists(file):
                    continue

                print(f'Processing : {file}')

                try:
                    if os.path.basename(file) in dataset:
                        u.append(dataset[os.path.basename(file)])
                    else:
                        u.append(preprocess(file))

                    new_input_files.append(file)
                except Exception as e:
                    print(e)
                    print(traceback.format_exc())
                    continue

            if len(u) > 0:
                u = numpy.stack(u, axis=0)
                scores = classifier.predict_proba(u)[:, 1].tolist()
                for file, score in zip(new_input_files, scores):
                    if score < 0.01:
                        try:
                            shutil.move(file, os.path.join(false_dir, os.path.basename(file)))
                        except OSError:
                            continue
                    elif score > 0.99:
                        try:
                            shutil.move(file, os.path.join(true_dir, os.path.basename(file)))
                        except OSError:
                            continue

コードが汚い。