Year2012 – Ongoing
RoleDeep Learning, DCNN
TechnologyDCNN, OpenCV, Python Keras, Tesseract, OCR Engine
Fontli, a Social network for typoholics
Fontli is a typography community platform that identifies images without typographic content and marks them as spam. The requirement was to introduce a moderation process to sift out non-typographic images (such as selfies and scenery) that are irrelevant to the typographic community.
Deep learning algorithms help spot the typeface
We employed the Deep Convolutional Neural Network (DCNN) technique to classify typographic/ non-typographic images and OpenCV to get rid of selfies and group photos. Tesseract was used for detecting text within an image and Python Keras library (we used over 45,000 images), to train the network.
Neural network see the letters
Fontli could achieve 90% accuracy based on historical data—the accuracy will improve with time. There was also an overall reduction in operational cost by replacing manual validation with technology.