I agree. I would look at doing something end to end, that can be completed in time. Also focus on making the implementation clean and having clear documentation - increase the chance of continuing to work after the course project deadline. CNN's are a good starting point.
I would look at the Final projects from Stanford's CS229 Machine learning course for inspiration.
Another idea would be to implement a subset of results from a recent machine learning paper, something that you are interested in. You could always extend this implementation depending on time and interest. For eg. this is a detailed relevant paper with a lot of modular sub-steps: RAISR: Rapid and Accurate Image Super Resolution
Don't spend too much time on selecting the project to work on, treat this as a way to master skills you can use for better things ahead.