Introduction

I was studying face recognition to do various things with machine learning, but I made it without taking a break. If you put an image of a person, it will judge whether it is an animation or a photo. Since it is based on the face, it may be judged as a photo if you include a realistic picture. I think it can be used by a pc to determine whether it is really close to reality through a live-action realistic picture.

Things to prepare

python3.○ OpenCV for python version Cascade classifier for photographic animation (described later)

development of

It's a simple thing, so let's code it as soon as you get what you need.

Implementation

(Folder) img-Image storage
(Folder) haarcascade-Cascade classifier storage
(File) auto_anime_classifier.py

Please place the above items in the same directory The files to put in the haarcascade folder this time

haarcascade_frontalface_alt.xml
lbpcascade_animeface.xml These two. The former is included in OpenCV by default. The latter is https://github.com/nagadomi/lbpcascade_animeface

Source code

`auto_anime_classifier.py`


import cv2
import numpy as np

if __name__ == '__main__':

    CASCADE_PATH = 'haarcascade/'
    face_count = []

    print('Plz input filename')
    img_name = input('>> ')
    img = cv2.imread('img/'+str(img_name), cv2.IMREAD_COLOR)
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

    photo = cv2.CascadeClassifier(CASCADE_PATH + 'haarcascade_frontalface_alt.xml')
    anime = cv2.CascadeClassifier(CASCADE_PATH + 'lbpcascade_animeface.xml')

    face_count.append(len(photo.detectMultiScale(gray, 1.1, 3)))
    face_count.append(len(anime.detectMultiScale(gray, 1.1, 3)))

    if face_count[0] > face_count[1]:
        print('Maybe photo.')
    elif face_count[1] > face_count[0]:
        print('Maybe anime.')
        elif face_count[0] >= 1 and face_count[1] >= 1:
        print('Both attributes.')
    elif face_count[0] == 0 and face_count[1] == 0:
        print('Undeciable.')

Commentary

About OpenCV, I will omit most of it because there are a lot of commentary blogs with wonderful contents. Enter the image in the img folder and read it with imread (). After that, it is grayed.

print('Plz input fiｌename')
img_name = input('>> ')
img = cv2.imread('img/'+str(img_name), cv2.IMREAD_COLOR)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

Create a classifier that reads the cascade classification files for each photo and animation.

photo = cv2.CascadeClassifier(CASCADE_PATH + 'haarcascade_frontalface_alt.xml')
anime = cv2.CascadeClassifier(CASCADE_PATH + 'lbpcascade_animeface.xml')

Store the number read by each classifier in face_count. The presence or absence of this is the criterion.

face_count.append(len(photo.detectMultiScale(gray, 1.1, 3)))
face_count.append(len(anime.detectMultiScale(gray, 1.1, 3)))

Finally, the judgment part. Compare whether it is simply recognized or not.

if face_count[0] > face_count[1]:
   print('Maybe photo.')
elif face_count[1] > face_count[0]:
   print('Maybe anime.')
elif face_count[0] >= 1 and face_count[1] >= 1:
   print('Both attributes.')
elif face_count[0] == 0 and face_count[1] == 0:
   print('Undeciable.')

Closeness to reality Maybe photo (probably a live-action film) → Both attributes (either one works) → Maybe anime (maybe a picture) → Undeciable (I don't know.

play

Let's play the explanation moderately.

Use normally

The Beatles. It is cool~

Plz input filename
>> beatles.jpeg
Maybe photo.

It looks like a photo. That's right.

The main character of NEW GAME. cute~

Plz input filename
>> aoba.jpeg
Maybe anime.

It seems to be an anime. Well, you can see it if you look at it.

Play with subtle images

Even if I did it normally, nothing was interesting, so let's put in a video that I do not know which one the program shows.

This is Mr. Layer SeeU from China. The image is a quote from his Twitter account.

Plz input filename
>> seeu.png
Maybe anime.

It seems to be an anime. From a human perspective, it doesn't look like a picture. Does image processing mean anything?

It is a character created by MMORPG called Black Desert. A type of game that takes hours to characterize.

Source (ttp: //sumimarudan.blog7.fc2.com/blog-entry-2245.html) *

Plz input filename
>> kuroisabaku.png
Both attributes.

It seems that you can go either way. Except for the big black eyes and the small nose, I don't think it feels like a picture.

It is a male character made in a game called Honey Select. You can make illusion characters without any discomfort even for characters other than young people.

Plz input filename
>> honey.png
Maybe photo.

It seems to be a photo. I wonder if the shape of the nose is an important turning point. By the way, the male character made with PSO2 was treated as a picture.

A doll from Orient Industry. It's very real.

Plz input filename
>> orient.jpg
Both attributes.

There was an anime element. I wonder why ...

At the end

The program this time was just about whether each judge could recognize the face. If it were to be improved, I felt that it was necessary to output detailed judgment elements from each judgment device and create parameters as numerical values. You have to study cascading classification. It's been a long time, but thank you for watching so far! !!

A program that automatically determines whether an animation or a photo is entered when a person's image is input [python]