The problem Safe Internet Usage solves

The internet is a wonderful place. While we can leverage its potential to gain information about anything and everything, it also brings a lot of potential risks. When we are exploring the web, we are vulnerable. Especially, the children these days are quickly hooked with the internet and more often stumble across content which is inappropriate for their age group. Such contents include inaccurate information or images that might upset them, and lead or tempt them into unlawful or dangerous behavior.
In the current scenario, most of the videos and abusive content are flagged by users that find it offensive. The content is then manually checked and appropriate action is taken such as removing the content or banning the user altogether. This system is very inefficient as it includes extreme time consumption and manual labor.
Our project aims to alleviate the manual labor needed for flagging safe/unsafe videos through the use of an intelligent system that does the process automatically and accurately.
Under the given system, the maximum focus is given to Real-time analysis of videos for its classification of into Safe/ Not Safe for children. This classification is done on the basis of the violent content of the video.
From the perspective of the various challenges faced in today’s scenario, our solution accomplishes the challenging task of labeling internet content as offensive or safe. It is automatic, requiring little to no human intervention and is real-time based, providing accurate results immediately. It is multi-faceted and can be used in a variety of industries and security sectors, in detecting anomalies and taking appropriate legal actions such as informing the police of a shoplifting incident or alerting the driver of a huge pothole that he might encounter in his route.

Challenges we ran into

Anomaly Detection is a burning topic today and there are many challenges faced while solving a real-world problem concerning it. Throughout our project, we faced many challenges and yet handled them.
First of all, it was very difficult to manage a good annotated huge dataset to train our models. We managed to have a licensed dataset for VSD from InterDigital.
Another challenge that we faced was while training our models. We tried and tested various models that include:

• InceptionV3
• VGG16
  *Xception
• Resnet50
  After so many failed attempts, we succeeded with the ResNet152 followed by Convolutional LSTM2D.
  Deep Neural Network Models require a huge amount of time to get trained which we found quite limited for the hackathon.