Some approaches for detecting multi lingual and multi text from natural scene images

Abstract

Content in the text format helps to communicate the relevant and specific newlineinformation to users meticulously. Text detection in natural scene images finds newlineseveral applications in computer vision systems such as reading license plates, newlinedetecting street signs, retrieving images, performing mobile visual searches, newlineidentifying image-based geolocation, and assisting visually impaired persons. newlineText, an important way of communication, provides significant information for newlineannotation, indexing, and image structuring. During communication, text can be newlineconveniently embedded in scenes or documents and text information in images is newlineeasily perceivable by everyone. But text extraction is a challenging problem due newlineto the inconsistency in text size, style, color, orientation, and alignment. Also, newlineimages with low contrast, blur, noise, varying illumination, complex background, newlineand multi-lingual environment add to this difficulty. This research aims to newlineovercome the existing difficulty in scene text extraction by developing approaches newlinefor detecting text from natural scene images for various quadrilateral-type and newlinepolygon-type datasets. newlineThe first work uses amended maximally stable extremal region (a-MSER) newlinetogether with deep learning framework, You Only Look Once (YOLOv2) network. newlineThe a-MSER method is used to identify the region of interest based on the newlinevariation of MSER. This algorithm considers intensity changes between text and newlinebackground very effectively.The drawback of original YOLOv2, the poor detection newlinerate for small-sized objects, is overcome by employing a 1 × 1 layer with image newlinesize enhanced from 13 × 13 to 26 × 26. Focal loss is applied to improve newlineupon the existing cross entropy classification loss of YOLOv2 newline