Inspiration
AI has been a tool to try and solve many of the world's problems. From medical diagnoses to making the roads safer with self-driving cars. Today, we try and tackle one of the biggest challenges in the modern era, making quality memes or image macros.
What it does
Generates text for a given image macro. Uses machine learning to determine the good text for memes.
Example
How we built it
- Use webscraping to build a large set of memes.
- Group them by template using the image similarity.
- Use OCR to read the text.
- For each template, train a neural network to generate new text.
- Utilize google cloud virtual machine for faster computation with a lot of input memes. ??? profit
Challenges we ran into
- Being able to read the font. OCR does not like reading memes.
- Tuning the neural network to output text.
- Google Cloud GPU and CPU restrictions
- Disk quotas preventing us from running the code in Linux
- Python dependency issues for importing TensorFlow and similar libraries
Accomplishments that we're proud of
Using selenium to scrap lots of memes from the internet, Sorting memes by image format, Using google cloud to run and train the AI, Using tesseract and how to train and use OCR,
What we learned
How to use OCR,
How train a AI using text,
What's next for Deep Meme
Making OCR more accurate
Creating more formats
Social media bot Increase funniness
Built With
- google-cloud
- ocr
- python
- selenium
- tensorflow
- tesseract
Log in or sign up for Devpost to join the conversation.