Kannada Word Detection in Heterogeneous Scene Images

Text extraction from scene images has started gaining a lot of traction in recent years in the computer vision field as its applications is manifold. One of its sub-categories is scene text detection. Factors like complex backgrounds, curvature, orientation, image quality, and various font styles and sizes makes it a difficult task. Moreover, building a single language scene text detector in a multilingual setting, especially Indian scripts, is more challenging in contrast to a general text detector. In this Work We attempt the task of Kannada text localization as Well as Word detection in natural scenes, Which is underexplored in comparison to other Well-known Indian languages. To achieve this, We fine tune a You Only Look Once (YOLOv4) based object detection model on our Kannada scene text images dataset collected. Experiments on local boards and sign images captured on smartphone, showed a mean average precision of 93.5% and per image average detection speed of 30 milli-seconds; pointing towards real time use and integration With mobile applications.