ディープラーニングで顔写真から巨乳かどうかを判別してみる (うまくいったか微妙).md

#!/bin/bash #画像ゲッター #引数で人名を与えると、temp_ほげほげ_gazou/gazou_00001.jpg等という連番で画像を保存してくれる。 # getter "ほげほげ" function getter() { for name in "$@"; do echo -e '\n\n\n\n' targettext_encoded=`echo "$name" | perl -MURI::Escape -lne 'print uri_escape($_)'` url='https://www.bing.com/images/search?FORM=HDRSC2&q='${targettext_encoded} echo target is "${name}" "${url}" if [ -d temp_${name}_gazou ]; then echo already downloaded temp_${name}_gazou continue fi mkdir temp_${name}_gazou pushd temp_${name}_gazou wget -r -l 1 -H "${url}" rm -rf tse1.mm.bing.net/ tse2.mm.bing.net/ tse3.mm.bing.net/ tse4.mm.bing.net/ www.bing.com/ www.bingfudosan.jp/ find ./ -not -name '*.jpg' -a -not -name '*.jpeg' -a -not -name '*.JPG' -type f -print0 | xargs -0 rm find ./ -type f | awk '{printf("mv '"'"'%s'"'"' gazou_%05d.jpg\n", $0, NR);}' > ../temp_${name}_gazou_filename_henkan.sh bash ../temp_${name}_gazou_filename_henkan.sh find ./ -maxdepth 1 -type d -print0 | xargs -0 rm -rf popd #連続でwgetするのを防ぐため。 sleep 5 done } getter "ほげほげ子" "もげもげ美" ・・・

//実行方法は以下。 // hoge.exe ターゲットのファイル名出力ファイル名 cascadeファイル名 //戻り値は、0が正常終了、100より下は引数不足等の実行時の異常、100以上はOpenCVは走ったが検出出来なかった等 #include "stdafx.h" #include <opencv2/opencv.hpp> #include <iostream> #include <windows.h> using namespace std; using namespace cv; #define FACE_KAKUDAI_WARIAI 1.2 //例えば検出された顔が100px四方だとしたら、100×FACE_KAKUDAI_WARIAIpxのサイズで切り抜く int main(int argc, char const *argv[]) { if(argc!=4){ printf("error. too few argument.\n"); exit(1); } printf("target file is %s\n", argv[1]); Mat image = imread(argv[1]); if(image.data==NULL){ printf("error. can not read source picture file.\n"); exit(2); } printf("image size width=%d height=%d\n", image.size().width, image.size().height); string cascade_filename = argv[3]; CascadeClassifier cascade; if(cascade.load(cascade_filename)!=true){ printf("error. can not read cascade file.\n"); exit(3); } vector<Rect> faces; cascade.detectMultiScale(image, faces, 1.1, 3, 0); printf("detect %zu faces.\n", faces.size()); if(faces.size()!=1){ printf("too many faces or no faces detected. exit.\n"); exit(100); } for (int i = 0; i < faces.size(); i++) { printf("face[%d] x=%d y=%d width=%d height=%d\n", i, faces[i].x, faces[i].y, faces[i].width, faces[i].height); //切り抜くサイズを少し大きくした時に画像をはみ出るかどうか確認。 int kakudai_size = (int)(faces[i].width * (FACE_KAKUDAI_WARIAI-1) / 2); //左右で拡大するので2で割っている。 int kirinuki_x = faces[i].x-kakudai_size; int kirinuki_y = faces[i].y-kakudai_size; int kirinuki_width = faces[i].width + kakudai_size*2; int kirinuki_height = faces[i].height+ kakudai_size*2; if( kirinuki_x<0 || kirinuki_y<0 || image.size().width <= kirinuki_x+kirinuki_width || image.size().height <= kirinuki_y+kirinuki_height ){ printf("face[%d] located on corner. skip.\n", i); exit(101); } Mat cut_img(image, Rect(kirinuki_x, kirinuki_y, kirinuki_width, kirinuki_height)); imwrite(argv[2], cut_img); printf("face[%d] outputed as \"%s\"\n", i, argv[2]); rectangle(image, Point(faces[i].x, faces[i].y), Point(faces[i].x + faces[i].width, faces[i].y + faces[i].height), Scalar(0, 200, 0), 1, CV_AA); rectangle(image, Point(kirinuki_x, kirinuki_y), Point(kirinuki_x + kirinuki_width, kirinuki_y + kirinuki_height), Scalar(200, 0, 0), 1, CV_AA); } //imshow("detect face", image); //waitKey(0); return 0; }

#!/bin/bash #顔部分だけ切り抜く。 #「./faces/temp_*_gazou/」以下の各gazou_*.jpgに対して処理を実行。 #「./faces/temp_*_gazou/face_gazou_00001.jpg」等のファイルができる。 CASCADE_FILENAME='(haarcascade_frontalface_alt2.xmlへのパス)' for i in ./faces/temp_*_gazou; do echo -e '\n\n\n\n' echo target dir $i pushd $i for j in gazou_*.jpg; do echo -e '\n' echo -e 'target dir is' "$i" ' target file is' "$j" if [ -f face_$j ]; then echo 'already detected face. skip' continue fi echo target file $j ./face_kirinuki.exe $j face_$j ${CASCADE_FILENAME} #もし判別に失敗してface_*.jpgのファイルができなかった場合、空のファイルを作る。 if [ $? -ne 0 ]; then touch face_$j continue fi #サイズを同じにする mogrify -geometry 256x256 face_$j done popd done

$ ./face_kirinuki.sh

$ sudo apt-get install python-pip python-dev $ sudo pip install --upgrade https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-0.6.0-cp27-none-linux_x86_64.whl

(ファイルパス) (分類巨乳なら1、貧乳なら0) ./faces/face_hogeko_00001.jpg 1 ./faces/face_mogemi_00001.jpg 0 ・・・

#!/usr/bin/env python # -*- coding: utf-8 -*- import sys import cv2 import numpy as np import tensorflow as tf import tensorflow.python.platform NUM_CLASSES = 2 IMAGE_SIZE = 28 IMAGE_PIXELS = IMAGE_SIZE*IMAGE_SIZE*3 flags = tf.app.flags FLAGS = flags.FLAGS flags.DEFINE_string('train', 'face_filelist_train.txt', 'File name of train data') flags.DEFINE_string('test', 'face_filelist_test.txt', 'File name of train data') flags.DEFINE_string('train_dir', './logdata', 'Directory to put the training data.') flags.DEFINE_integer('max_steps', 100, 'Number of steps to run trainer.') flags.DEFINE_integer('batch_size', 10, 'Batch size' 'Must divide evenly into the dataset sizes.') flags.DEFINE_float('learning_rate', 1e-4, 'Initial learning rate.') def inference(images_placeholder, keep_prob): """ 予測モデルを作成する関数引数: images_placeholder: 画像のplaceholder keep_prob: dropout率のplace_holder 返り値: y_conv: 各クラスの確率(のようなもの) """ # 重みを標準偏差0.1の正規分布で初期化 def weight_variable(shape): initial = tf.truncated_normal(shape, stddev=0.1) return tf.Variable(initial) # バイアスを標準偏差0.1の正規分布で初期化 def bias_variable(shape): initial = tf.constant(0.1, shape=shape) return tf.Variable(initial) # 畳み込み層の作成 def conv2d(x, W): return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') # プーリング層の作成 def max_pool_2x2(x): return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') # 入力を28x28x3に変形 x_image = tf.reshape(images_placeholder, [-1, 28, 28, 3]) # 畳み込み層1の作成 with tf.name_scope('conv1') as scope: W_conv1 = weight_variable([5, 5, 3, 32]) b_conv1 = bias_variable([32]) h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) # プーリング層1の作成 with tf.name_scope('pool1') as scope: h_pool1 = max_pool_2x2(h_conv1) # 畳み込み層2の作成 with tf.name_scope('conv2') as scope: W_conv2 = weight_variable([5, 5, 32, 64]) b_conv2 = bias_variable([64]) h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) # プーリング層2の作成 with tf.name_scope('pool2') as scope: h_pool2 = max_pool_2x2(h_conv2) # 全結合層1の作成 with tf.name_scope('fc1') as scope: W_fc1 = weight_variable([7*7*64, 1024]) b_fc1 = bias_variable([1024]) h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64]) h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1) # dropoutの設定 h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) # 全結合層2の作成 with tf.name_scope('fc2') as scope: W_fc2 = weight_variable([1024, NUM_CLASSES]) b_fc2 = bias_variable([NUM_CLASSES]) # ソフトマックス関数による正規化 with tf.name_scope('softmax') as scope: y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2) # 各ラベルの確率のようなものを返す return y_conv def loss(logits, labels): """ lossを計算する関数引数: logits: ロジットのtensor, float - [batch_size, NUM_CLASSES] labels: ラベルのtensor, int32 - [batch_size, NUM_CLASSES] 返り値: cross_entropy: 交差エントロピーのtensor, float """ # 交差エントロピーの計算 #cross_entropy = -tf.reduce_sum(labels*tf.log(logits)) 異常処理 http://qiita.com/ikki8412/items/3846697668fc37e3b7e0 cross_entropy = -tf.reduce_sum(labels*tf.log(tf.clip_by_value(logits,1e-10,1.0))) # TensorBoardで表示するよう指定 tf.scalar_summary("cross_entropy", cross_entropy) return cross_entropy def training(loss, learning_rate): """ 訓練のOpを定義する関数引数: loss: 損失のtensor, loss()の結果 learning_rate: 学習係数返り値: train_step: 訓練のOp """ train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss) return train_step def accuracy(logits, labels): """ 正解率(accuracy)を計算する関数引数: logits: inference()の結果 labels: ラベルのtensor, int32 - [batch_size, NUM_CLASSES] 返り値: accuracy: 正解率(float) """ correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) tf.scalar_summary("accuracy", accuracy) return accuracy if __name__ == '__main__': # ファイルを開く f = open(FLAGS.train, 'r') # データを入れる配列 train_image = [] train_label = [] for line in f: # 改行を除いてスペース区切りにする line = line.rstrip() l = line.split() # データを読み込んで28x28に縮小 print "aaaaa line[%s] %s %s"%(line,l[0],l[1]) img = cv2.imread(l[0]) img = cv2.resize(img, (28, 28)) # 一列にした後、0-1のfloat値にする train_image.append(img.flatten().astype(np.float32)/255.0) # ラベルを1-of-k方式で用意する tmp = np.zeros(NUM_CLASSES) tmp[int(l[1])] = 1 train_label.append(tmp) # numpy形式に変換 train_image = np.asarray(train_image) train_label = np.asarray(train_label) f.close() f = open(FLAGS.test, 'r') test_image = [] test_label = [] for line in f: line = line.rstrip() l = line.split() img = cv2.imread(l[0]) img = cv2.resize(img, (28, 28)) print "aaaaa2 line[%s] %s %s"%(line,l[0],l[1]) test_image.append(img.flatten().astype(np.float32)/255.0) tmp = np.zeros(NUM_CLASSES) tmp[int(l[1])] = 1 test_label.append(tmp) test_image = np.asarray(test_image) test_label = np.asarray(test_label) f.close() with tf.Graph().as_default(): # 画像を入れる仮のTensor images_placeholder = tf.placeholder("float", shape=(None, IMAGE_PIXELS)) # ラベルを入れる仮のTensor labels_placeholder = tf.placeholder("float", shape=(None, NUM_CLASSES)) # dropout率を入れる仮のTensor keep_prob = tf.placeholder("float") # inference()を呼び出してモデルを作る logits = inference(images_placeholder, keep_prob) # loss()を呼び出して損失を計算 loss_value = loss(logits, labels_placeholder) # training()を呼び出して訓練 train_op = training(loss_value, FLAGS.learning_rate) # 精度の計算 acc = accuracy(logits, labels_placeholder) # 保存の準備 saver = tf.train.Saver() # Sessionの作成 sess = tf.Session() # 変数の初期化 sess.run(tf.initialize_all_variables()) # TensorBoardで表示する値の設定 summary_op = tf.merge_all_summaries() summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, sess.graph_def) # 訓練の実行 for step in range(FLAGS.max_steps): for i in range(len(train_image)/FLAGS.batch_size): # batch_size分の画像に対して訓練の実行 batch = FLAGS.batch_size*i # feed_dictでplaceholderに入れるデータを指定する sess.run(train_op, feed_dict={ images_placeholder: train_image[batch:batch+FLAGS.batch_size], labels_placeholder: train_label[batch:batch+FLAGS.batch_size], keep_prob: 0.5}) # 1 step終わるたびに精度を計算する train_accuracy = sess.run(acc, feed_dict={ images_placeholder: train_image, labels_placeholder: train_label, keep_prob: 1.0}) print "step %d, training accuracy %g"%(step, train_accuracy) # 1 step終わるたびにTensorBoardに表示する値を追加する summary_str = sess.run(summary_op, feed_dict={ images_placeholder: train_image, labels_placeholder: train_label, keep_prob: 1.0}) summary_writer.add_summary(summary_str, step) # 訓練が終了したらテストデータに対する精度を表示 print "test accuracy %g"%sess.run(acc, feed_dict={ images_placeholder: test_image, labels_placeholder: test_label, keep_prob: 1.0}) # 最終的なモデルを保存 save_path = saver.save(sess, "model.ckpt")

$ python gakusyuu.py

$ tensorboard --logdir=(ログファイルのあるディレクトリの絶対パス) $ tensorboard --logdir=/home/hoge/kyonyuu_hanbetu/logdata

#!/usr/bin/env python #! -*- coding: utf-8 -*- import sys import numpy as np import tensorflow as tf import cv2 NUM_CLASSES = 2 IMAGE_SIZE = 28 IMAGE_PIXELS = IMAGE_SIZE*IMAGE_SIZE*3 def inference(images_placeholder, keep_prob): """ モデルを作成する関数引数: images_placeholder: inputs()で作成した画像のplaceholder keep_prob: dropout率のplace_holder 返り値: cross_entropy: モデルの計算結果 """ def weight_variable(shape): initial = tf.truncated_normal(shape, stddev=0.1) return tf.Variable(initial) def bias_variable(shape): initial = tf.constant(0.1, shape=shape) return tf.Variable(initial) def conv2d(x, W): return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') def max_pool_2x2(x): return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') x_image = tf.reshape(images_placeholder, [-1, 28, 28, 3]) with tf.name_scope('conv1') as scope: W_conv1 = weight_variable([5, 5, 3, 32]) b_conv1 = bias_variable([32]) h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) with tf.name_scope('pool1') as scope: h_pool1 = max_pool_2x2(h_conv1) with tf.name_scope('conv2') as scope: W_conv2 = weight_variable([5, 5, 32, 64]) b_conv2 = bias_variable([64]) h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) with tf.name_scope('pool2') as scope: h_pool2 = max_pool_2x2(h_conv2) with tf.name_scope('fc1') as scope: W_fc1 = weight_variable([7*7*64, 1024]) b_fc1 = bias_variable([1024]) h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64]) h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1) h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) with tf.name_scope('fc2') as scope: W_fc2 = weight_variable([1024, NUM_CLASSES]) b_fc2 = bias_variable([NUM_CLASSES]) with tf.name_scope('softmax') as scope: y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2) return y_conv if __name__ == '__main__': test_image = [] filenames = [] for i in range(1, len(sys.argv)): img = cv2.imread(sys.argv[i]) img = cv2.resize(img, (28, 28)) test_image.append(img.flatten().astype(np.float32)/255.0) filenames.append(sys.argv[i]) test_image = np.asarray(test_image) images_placeholder = tf.placeholder("float", shape=(None, IMAGE_PIXELS)) labels_placeholder = tf.placeholder("float", shape=(None, NUM_CLASSES)) keep_prob = tf.placeholder("float") logits = inference(images_placeholder, keep_prob) sess = tf.InteractiveSession() saver = tf.train.Saver() sess.run(tf.initialize_all_variables()) saver.restore(sess, "model.ckpt") for i in range(len(test_image)): pred = np.argmax(logits.eval(feed_dict={ images_placeholder: [test_image[i]], keep_prob: 1.0 })[0]) pred2 = logits.eval(feed_dict={ images_placeholder: [test_image[i]], keep_prob: 1.0 })[0] print filenames[i],pred,"{0:10.8f}".format(pred2[0]),"{0:10.8f}".format(pred2[1])

#!/bin/bash python ./hanbetu__all_file.py "$@" >> log_hanbetu_`date +%Y%m%d_%H%M%S`.txt

$ ./hanbetu_all.sh (判別したいファイル複数)

2016/08/23追記

2016/02/10追記

2016/03/14追記

前提

巨乳の方と貧乳の方をリストアップ

写真を片っ端から集める

写真から顔部分だけを切り抜き

顔から特徴抽出し学習

TensorFlowセットアップ

学習

学習したモデルを使って判定

まとめ