AdvancedEAST

Why we included this project

AdvancedEAST builds on the EAST scene text detector but changes how the four corners of each text box are predicted. Instead of averaging predicted coordinates across every pixel in a text region, the network isolates boundary pixels and lets their weighted predictions fix the corners, which helps on long or skewed lines of text where plain EAST tends to blur the far edge of a quadrangle. The whole thing is a compact Keras codebase split into small files for labeling, the network, loss, training, and prediction, so it reads well as a study aid for anyone learning how a text detector is assembled and trained end to end. It ships training and prediction entry points plus demos built around the ICPR web-image dataset, so you can reproduce results fairly quickly. One caveat: it targets TensorFlow 1.x and Keras 2.x, so plan on porting it if you run a modern TensorFlow.