ethanstanks/violenceclassificationcv

Computer Vision classification for violence using a VGG16 Architecture with pictures and videos

Computer Vision Violence Classification

Description

Computer Vision classification for violence using a VGG16 Architecture with pictures and videos.

Features

• VGG16 Architecture: Three Convolution2D layers with a FC dense layer as the output
• Libraries: TensorFlow Keras and Sklearn for the model. CV2 and OS for data preprocessing.
• Image Prediction: Predicts violence for multiple images. Graphs their predictions with the original images using MatPlotLib.
• Video Prediction: Predicts violence for each frame of a video.
• Video Blurring: Capable of blurring violence detected frames of videos then writing the altered version to a new .mp4v

Model's Accuracy & Loss

Image Prediction Example

Video Violence Blurring Example (click image)

Installation and Running the Script

To install and run the model on your Windows machine, follow these steps:

1. Clone the Repository: Download the codebase to your local machine.
   https://github.com/EthanStanks/ViolenceClassificationCV.git
2. Install Dependencies: Ensure you have Python installed, then use pip to install the required libraries.
   pip install -r requirements.txt
3. Download Dataset: Download the dataset from Kaggle then place the images inside:.
   data/train_images/
4. Train Model: Set TRAIN_MODEL boolean to True. Navigate to the src directory and execute the main.py script.
   python main.py
5. Predict Images Add images you'd like to predict inside data/validate_images/. Set PREDICT_IMAGES boolean to True. Add image paths to list on line 95. Execute the main.py script.
   PREDICT_IMAGES = True
6. Predict Videos Add the video you'd like to predict inside data/video_input/. Set PREDICT_VIDEO boolean to True. Add the file name to the function call on line 88. Execute the main.py script.
   PREDICT_VIDEO = True
7. Blur Videos Follow step 6. Ensure True is passed as second parameter on line 88. Execute the main.py script.
   PREDICT_VIDEO = True

Credits

Data for this project is sourced from Kaggle's "Violence vs. Non-Violence: 11K Images Dataset" dataset, available here.

---

Note: This project is currently tailored for Windows operating systems.