Detecting and recognizing text in a natural setting is a challenging task due to various shapes and unpredictable environment these texts are a part of. Most methods have successfully detected text instances in an environment despite the environment’s complexity. However, most methods assume the text instances are somewhat straight or in a linear format. However, it is quite common to find text that is “curved” or non-linear in the real world. Therefore, applying a text recognition model on the output of a text detection model using STN fails to achieve desired results. Furthermore, even after detecting the text, one must rectify the text before feeding it into the text recognition model. Therefore, I propose a method that combines a previous text detection model((Long et al., 2018)) with a text rectification algorithm: Thin Plate Spline (TPS) and a text recognition model((Bartz et al., 2018)) to identify curved text in natural settings.
Paper can be found here