| Summary: | Scene text detection methods in computer vision and object detection relying heavily on neural network and deep learning have emerged recently, showing promising results. The topic of object recognition is a subject of ongoing active research and have been used in applications such as text localization, surveillance, aerial imaging and autonomous driving. In addition, applications of scene text detection include multilingual text translation on mobile phone aiding users with instant translation, blind-navigation and image information retrieval.While recent advancement in the field have led to improved accuracies, precisions and f-measures, text extraction from natural scenes still pose as a challenging problem often involve with complex issues. These problems include, dis-oriented text, perspective distortion, arbitrary shaped text, variation in text sizes, uneven lighting and blurring. State-of-the-art object detector localize text of interest accurately by drawing horizontal/vertical, rectangular shaped bounding boxes over an object. These methods fail to address the issue of perspective distortion, variation in text sizes and arbitrary shaped text resulting in reduced text localization and limiting downstream task. To overcome these limitations, this research paper introduces two solutions that addresses these issues to achieve more precise detections producing better fitted and tighter bounding boxes. Specifically,by adding a new parameter call angle in the anchor box parameter definition and predict a rotation of clockwise or anti-clockwise around it’s midpoint and by allowing the anchor boxes to morph around it four point definitions. As a result, further improving the text localization network accuracy with minimum impact in its reliability and speed.
|
|---|