Full Submission

Full submission is linked on GitHub.

Inspiration

The detection and identification of TV station logos can be used to improve the user experience as it allows to remove commercial breakes from recorded TV shows and movies. In addition, it can be used to detect product piracy in the Internet and forbidden symbols in video frames. Platforms like YouTube could use such algorithms to automatically remove videos containing forbidden symbols or illegally uploaded videos.

What it does

UnCannyEdges is used to identify and classify the logos of television channels on video frames. Therefore, we used machine learning techniques to generate a model from a large set of training samples. Those samples contain the logo and were classified by hand by Rohde & Schwarz. Our algorithm uses this training data to learn 'how the logos look like' and is able to identify them on unknown frames.

How we built it

We used our own Python scripts (utilizing Numpy) to implement basic image filtering algorithms, like noise removal, Canny edge detection. In addition, we used the Tensorflow framework to build a neuronal network trained to detect the logos.

Challenges we ran into

The challenge is split in two parts. On the one hand we have to identify a logo on the different video frames. On the other hand, you have to detect the position of the logo on the video frame. Our opinion is that the second task is the more complex and resource critical part of the task.

Accomplishments that we're proud of

We are proud that we tried many different techniques to detect and classify the logos, like SURF, Canny, adn ORB algorithms. In addition, we spent a lot of time in brain storming how we could solve the task and read a lot of articles about object detection to get a well suited and self-built solution instead of taking a ready-to-use framework. A lot of efford was the data preprocessing to get optimal data for the neuronal network training.

What we learned

In general, we learned a lot of image processing in Python, like filtering. Learned a lot about the state of the art in object detection and machine learning techniques in general. It was very interesting to see how powerful neuronal network are. In addition, we had an introduction in the use of the powerful machine learning framework Tensaflow.

What's next for UnCannyEdges

Get some sleep... Use more training data and parameter tuning to optimize the neuronal network. In addition, we want to understand neuronal network in detail to be able to use its various capabilities.

Built With

Share this project:
×

Updates