{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# 主題：健康上網AI守門員！ 色情圖片辨識\n", "### 組員：陳怡升、張睿傑、李志恆
\n", "[發表簡報請點我](https://docs.google.com/presentation/d/19mwSTLK1vLj8kBd4NtedvkfHCA17YikbTqA5CD2D2Tg/edit?usp=sharing)\n", "
\n", "### 概述：\n", "我們打造出一個新型態的健康上網守門員
透過色情圖片辨識，當偵測到使用者在瀏覽色情內容時，將會自動播放歌曲[Big Enough](https://www.youtube.com/watch?v=rvrZJ5C_Nwg)中[牛仔的尖叫片段](https://www.youtube.com/watch?v=Qcp2W1-SFt4)，
然後讓滑鼠亂飄移，並按下Alt+F4關閉視窗讓使用者無法正常瀏覽色情內容。\n", "![AHHH](https://trello-attachments.s3.amazonaws.com/5c6a07f61246780dd23d1ee9/5cf9e77c1b3c9662de99a4de/9d3406ef219953d95918ff4721942cf8/image.png)\n", "
程式實際使用示範影片：[有碼版本示範影片](https://www.youtube.com/watch?v=AsDYYk-qPA8&feature=youtu.be)\n", "
同場加映：[上傳到PornHub的無碼版本示範影片](https://www.pornhub.com/view_video.php?viewkey=ph5cfbb88976949) (內含18+畫面，慎入！！)\n", "\n", "\n", "### 目的：\n", "本來我們想做色情圖片辨識，不過想想辨識出來也要有個用途，靈光一閃就想到這個\"健康上網AI守門員\"的idea
\n", "我們有空的話，會在暑假將其改良成開機後自動在背景執行，為我國青少年身心健康發展盡一份心力。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 具體做法：\n", "* **資料取得：**
\n", "我們將透過github上的[NSFW數據庫](https://github.com/GantMan/nsfw_model)來取得資料進行訓練。
\n", "將資料集中corrupted image 移除後剩下的**212038**張圖片被分為5種類別，分別為：
\n", "![CLASSES](https://trello-attachments.s3.amazonaws.com/5c6a07f61246780dd23d1ee9/5cf9e77c1b3c9662de99a4de/a695856e957d053665144284c3ca3dd4/%E4%B8%8B%E8%BC%89.png)\n", "

\n", "* **模型訓練：**
\n", "使用Keras中的InceptionV3模型中ImageNet 的權重來進行transfer learning，再加上兩層DenseNet提高辨識度。

\n", "* **訓練結果：**
\n", "在經過長時間訓練後，模型實際使用時正確率很高，在將圖片分類上能夠達到90% 以上的正確率。

\n", "訓練模型建置的github 連結: https://github.com/juichiehchang/NCTS5017-PythonFinalProject

\n", "* **程式建置：**
\n", "我們把訓練完成的**模型**與**權重**讀入，打造出一個能夠辨識輸入圖片類別的辨識器。
\n", "接著使用PIL套件截圖，並將圖片分割為18個小圖片進行辨識，如下圖
\n", "![CLASSES](https://trello-attachments.s3.amazonaws.com/5c6a07f61246780dd23d1ee9/5ce74f673f93f75eff62e9b1/ac0330eee3fa696814a4df43863c9331/image.png)\n", "(我們將螢幕分割為3x5共15張圖片，並在影片播放位置的**重點區域**多剪兩張以及整個畫面一張，總共18張)

\n", "該程式將定期做螢幕截圖並辨識圖片中是否有色情內容。
\n", "如果色情內容的比例超過一定水準，將會使用winsound套件自動播放歌曲Big Enough中牛仔的尖叫片段，
\n", "並使用PyAutoGUI讓滑鼠隨機飄移無法操控並在十五秒後按下ALT+F4關閉視窗
\n", "而如果判斷色情內容比例超過一定水準，但尚未達到該關閉的程度，
\n", "因為可能會有誤判也會達到該水平，故程式會播放「光頭哥哥的嘿嘿」音效來提醒使用者，並將記錄累加，如持續超過一定水準也會執行關閉程式。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 此程式的缺點：\n", "使用本模型，有一個缺點，就是**非常耗CPU資源**，因為判讀模型很大
\n", "當執行判斷時，我們的CPU使用率會衝到100%約3秒，如下方圖片所示。
\n", "(高峰為程式執行判斷的時候)
\n", "![IMAGES](https://trello-attachments.s3.amazonaws.com/5c6a07f61246780dd23d1ee9/5cf9e77c1b3c9662de99a4de/9638d71ae5f2051ebe93f0c0a96d4c99/%E6%93%B7%E5%8F%96.JPG)\n", "未來如果要改進，需要尋找有無抑制CPU使用率的寫法，
\n", "或將截圖上傳至雲端，由雲端的伺服器做判斷後，再將結果回傳給用戶端執行。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# 程式碼內容\n", "(歡迎複製過去自行使用，其他檔案請至[雲端硬碟下載](https://drive.google.com/file/d/1nRXHMx9T4MLmZBP2rfjc8G8krewEHhaD/view?usp=sharing)整合好的zip檔案)
後面還有趣事(判斷失敗的例子)喔！" ] }, { "cell_type": "code", "execution_count": 2, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "C:\\Users\\Public\\ANACONDA3\\lib\\site-packages\\h5py\\__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n", " from ._conv import register_converters as _register_converters\n", "Using TensorFlow backend.\n" ] } ], "source": [ "#import predict\n", "import cv2\n", "import numpy as np\n", "from PIL import Image\n", "from PIL import ImageGrab\n", "import time\n", "import pyautogui\n", "import winsound\n", "#import sys\n", "#sys.setrecursionlimit(50)\n", "import numpy as np\n", "import os\n", "os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'\n", "import keras\n", "\n", "def load_image(img_path, size):\n", " image = keras.preprocessing.image.load_img(img_path, target_size = size)\n", " image = keras.preprocessing.image.img_to_array(image)\n", " image /= 255\n", " return np.expand_dims(image, axis=0)\n", "\n", "class predictor():\n", "\n", " model = None\n", " categories = ['drawings', 'hentai', 'neutral', 'porn', 'sexy']\n", " size = None\n", " \n", " def __init__(self, model_path):\n", " if('inceptionV3' in model_path):\n", " self.size = (299, 299)\n", " self.model = keras.models.load_model(model_path)\n", " \n", " def predict_from_path(self, img_path):\n", " \n", " image = load_image(img_path, self.size)\n", " prediction = self.model.predict(image)\n", " #print(prediction)\n", " return self.categories[np.argmax(prediction)]\n", "\n", " def predict_from_array(self, img):\n", "\n", " image = img/255\n", " prediction = self.model.predict(np.expand_dims(image, axis=0))\n", " return self.categories[np.argmax(prediction)]\n" ] }, { "cell_type": "code", "execution_count": 4, "metadata": { "scrolled": false }, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "C:\\Users\\Public\\ANACONDA3\\lib\\site-packages\\keras\\engine\\saving.py:327: UserWarning: Error in loading the saved optimizer state. As a result, your model is starting with a freshly initialized optimizer.\n", " warnings.warn('Error in loading the saved optimizer '\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "Judge start\n", "讚喔\n", "1\n", "drawings: 10\n", "hentai: 0\n", "neutral: 6\n", "porn: 0\n", "sexy: 2\n", "Judge end\n", "\n", "Judge start\n", "母湯喔\n", "0\n", "drawings: 1\n", "hentai: 0\n", "neutral: 7\n", "porn: 10\n", "sexy: 0\n", "Judge end\n", "\n", "Judge start\n", "讚喔\n", "1\n", "drawings: 1\n", "hentai: 0\n", "neutral: 17\n", "porn: 0\n", "sexy: 0\n", "Judge end\n", "\n", "Judge start\n", "讚喔\n", "1\n", "drawings: 9\n", "hentai: 0\n", "neutral: 9\n", "porn: 0\n", "sexy: 0\n", "Judge end\n", "\n", "Judge start\n", "讚喔\n", "1\n", "drawings: 12\n", "hentai: 0\n", "neutral: 6\n", "porn: 0\n", "sexy: 0\n", "Judge end\n", "\n", "Judge start\n", "讚喔\n", "1\n", "drawings: 7\n", "hentai: 0\n", "neutral: 10\n", "porn: 1\n", "sexy: 0\n", "Judge end\n", "\n", "Judge start\n", "讚喔\n", "1\n", "drawings: 2\n", "hentai: 0\n", "neutral: 15\n", "porn: 1\n", "sexy: 0\n", "Judge end\n", "\n", "Judge start\n", "讚喔\n", "1\n", "drawings: 5\n", "hentai: 0\n", "neutral: 13\n", "porn: 0\n", "sexy: 0\n", "Judge end\n", "\n" ] } ], "source": [ "weights_path = \"C:/Users/user/Downloads/pythonF/inceptionV3.299x299.h5\" #請自行修改路徑\n", "\n", "p = predictor(weights_path)\n", "categories = ['drawings', 'hentai', 'neutral', 'porn', 'sexy']\n", " \n", "#截圖程式\n", "#自動在螢幕產生15個截圖\n", "def GrabAndCut():\n", " images = []\n", " width = 300 \n", " height = 300\n", " im=ImageGrab.grab() #截圖\n", " wid_1 = -300\n", " hei_1 = -300\n", " for i in range(3):\n", " wid_1 = -300\n", " for i in range(5):\n", " im_temp = im.crop(( #裁切圖片，裁切出3*5個\n", " width +wid_1,\n", " height + hei_1,\n", " width +wid_1 + 300,\n", " height + hei_1 + 300))\n", " images.append(im_temp.resize((299,299)))\n", " wid_1+=300\n", " hei_1+=300 \n", " images.append(im.crop(( #多剪兩張在平常的播放位置的內容\n", " 100,\n", " 200,\n", " 800,\n", " 800)).resize((299,299))) #將圖片大小設定為299x299\n", " images.append(im.crop((\n", " 500,\n", " 200,\n", " 1100,\n", " 800)).resize((299,299)))\n", " images.append(im.resize((299,299)))\n", " return images\n", "\n", "def CheckImages(images): #分類函式 將圖片分類\n", " result = []\n", " for i in range(len(images)):\n", " result.append(p.predict_from_array(np.array(images[i])))\n", " return result\n", " #判斷時要使用np.array判斷，所以我們將圖片轉換成array\n", " #然後將判斷結果連接到一個叫result的陣列\n", "\n", "\n", "def Judge(result, beforehand): #根據結果做出判斷的程式\n", " testing = result.count(\"porn\")*2 + result.count(\"hentai\")\n", " #使用自訂的判斷權重，因為在看動漫時較容易判斷出hentai故hentai需要比較高的權重\n", " if testing >=9 or (testing + beforehand) >=11:\n", " winsound.PlaySound(\"AHHH.wav\", winsound.SND_FILENAME|winsound.SND_ASYNC)\n", " #使用winsound播放音樂， winsound.SND_FILENAME 是指播放該檔名的檔案\n", " #winsound.SND_ASYNC 是指播放後就繼續執行剩下的程式碼\n", " for i in range(40): #亂動滑鼠的部分\n", " pyautogui.moveTo(np.random.randint(1,1910), np.random.randint(1,1070),0.25)\n", " pyautogui.keyDown(\"ALT\") #使用PyAutoGUI的套件，模擬按下ALT+F4\n", " pyautogui.keyDown(\"F4\")\n", " pyautogui.keyUp(\"ALT\")\n", " #x, y =pyautogui.size() #移動到右上角關閉視窗\n", " #pyautogui.moveTo(x-10, 10, 10) #後來選擇ALT+F4，故此段註解\n", " #pyautogui.click() \n", " print(\"母湯喔\")\n", " Judges = 0\n", " elif testing >=5:\n", " print(\"嘿嘿\") #播放光頭哥哥的「嘿嘿」音效，警告使用者\n", " winsound.PlaySound(\"hehe.wav\", winsound.SND_FILENAME|winsound.SND_ASYNC)\n", " Judges = 1\n", " beforehand = testing #將結果回傳，如果重複出現危險內容將會執行關閉程式\n", " else:\n", " print(\"讚喔\")\n", " Judges = 1\n", " beforehand = 0\n", " return Judges, beforehand\n", "\n", "beforehand = 0\n", "#while(1): #正式程式，將內容串在一起\n", "for i in range(8): #本來是使用while迴圈，不過因為要呈現不想出現KeyboardInterrupt\n", " #所以使用for迴圈做八次\n", " print(\"Judge start\")\n", " images = GrabAndCut()\n", " result = CheckImages(images)\n", " Judges, beforehand = Judge(result, beforehand)\n", " print(Judges)\n", " for i in range(len(categories)): #打印出分類模型究竟分類出來的結果\n", " print(\"%s: %d\"%(categories[i], result.count(categories[i])))\n", " print(\"Judge end\\n\")\n", " time.sleep(1)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 趣聞 (判斷失敗的例子) ：\n", "此程式準確率還是有待加強，比方會將一些新聞的圖片判斷為porn\n", "(如下方圖片)
\n", "推測可能是因為原先training data **並沒有文字**，同時切割過的圖片較不易判斷
所以我們使用指定公式，當porn分類超過5個，才執行關閉程式。\n", "![IMAGE](https://trello-attachments.s3.amazonaws.com/5c6a07f61246780dd23d1ee9/5cf9e77c1b3c9662de99a4de/74eadfc81e7e8158d61fb029b8b473e4/messageImage_1559742855638.jpg)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.7.3" }, "toc": { "base_numbering": 1, "nav_menu": {}, "number_sections": false, "sideBar": true, "skip_h1_title": false, "title_cell": "Table of Contents", "title_sidebar": "Contents", "toc_cell": false, "toc_position": { "height": "calc(100% - 180px)", "left": "10px", "top": "150px", "width": "165px" }, "toc_section_display": true, "toc_window_display": true } }, "nbformat": 4, "nbformat_minor": 2 }