A semi-automatic framework for text insertion and replacement in natural images
Abstract
Digital printing brings about a host of benefits, one of which is the ability to create short runs of variable, customized content. Specifically in photo-finishing and direct marketing applications, a natural image can be personalized by adding new text or changing existing text in an intelligent and natural way. Current solutions to achieve the aforementioned natural text insertion and replacement are however highly manual, requiring a lot of time, effort and skills. In this dissertation, a semi automatic framework is proposed, which is capable of inserting and replacing text on both planar and cylindrical surfaces. Techniques in image processing and computer vision such as the bilateral filter, the Hough transform, the Kullback-Leibler divergence, the polygon triangulation, and the connected component analysis are combined, along with necessary user inputs. Unlike current solutions which utilize 2-D text rendering, a 3-D pinhole camera is used as our rendering model, which introduces significant flexibility. Also, an image pre-analysis and recommendation algorithm is developed and included in the framework to help pre-screen a set of images according to their suitability for personalization, i.e., whether there are smooth regions and/or existing text inside the images. Popular learning-based classifiers such as multinomial logistic regression are utilized. The semi-automatic framework has been implemented as a tool in Java. The developed tool and experimental results are shown to demonstrate the effectiveness of our proposed framework.
Degree
Ph.D.
Advisors
Bouman, Purdue University.
Subject Area
Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.