Video Content Moderation using Neural Networks (Year 2)
2019-2020
Proposed by Peter
Project Description:
As web technology grows rapidly and mature, people can upload their personal images or videos to the Internet through any social media and video-sharing websites, such as YouTube and Facebook. However, some bad guys may use those platforms to upload their harmful or dangerous content such as violent events and bloody content to the Internet, those kinds of video content may be illegal. Nowadays, artificial intelligence can help us to solve this problem. People can utilize AI in the pre-moderation stage, and filter out most of the non-compliant videos by analyzing video content. This can save a lot of time and increase the moderation accuracy.
This project aims to analyze video content using deep neural networks. For simpler, the requirement of analyzing video content is to check if cats are present in a provided video and capture the moment cat apparent. Training dataset will be provided.
Keras
Keras is a high-level neural networks API, written in Python and capable of running on top of TensorFlow, CNTK, or Theano.
Main tasks:
In this project, a dataset for the training and a video for the testing are provided. The script to process videso is also provided. The requirements are listed below:
• Build Neural Networks based on Keras.
• Train the Neural Networks by using the given dataset. The given dataset should be split into two parts: one part is for training; another is for testing.
• Record log files during the training and monitor the training process using TensorBoard.
• After training, examine the trained model by two tests: develop a simple test program to examine the accuracy of the final trained model by detecting cats on non-trained images and videos with the provided scripts. The final well-trained model with the log file and the test program will to be submitted as the result.
• The evaluation accuracy on images and videos are the one of the important marking criterions, so try your best to design and adjust your model to get a better performance.
• Use the well-trained model for recognizing cats from all the frames of a video and save all the frames that cats appear as image files.
Bonus task:
Students can obtain extra marks by finishing bonus tasks and it is not mandatory.
• Students can develop a simple web crawler that is used to collect training images from the Internet, which may increase the training accuracy since there would be more training images.
• Students can develop a program to generate new video by using the recognized frames.
Useful Materials:
• Images dataset and Videos
https://bit.ly/2Tbn0iH
https://bit.ly/2SIGR8U
• Neural Network Tutorial
https://bit.ly/2ZGeB8h
• Web Crawler Tutorial
https://scotch.io/tutorials/getting-started-with-python-requests-get-requests
https://www.codementor.io/@aviaryan/downloading-files-from-urls-in-python-77q3bs0un
• Image Site
https://shutr.bz/2SJ9m6s