טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentGoferman Stas
SubjectPuzzle-like Collage
DepartmentDepartment of Electrical Engineering
Supervisors Professor Lihi Zelnik-Manor
Professor Ayellet Tal
Full Thesis textFull thesis text - English Version


Abstract

A collage is a work of the visual arts, made from an assemblage of different forms, thus creating a new whole, often having a purposeful incongruity. This work focuses on photo-collages, which assemble a collection of photographs by cutting and joining them together. A photo-collage can be used for art as well as for summarizing a photo collection, such as a news event, a family occasion, or a concept. Manually creating a collage is a difficult and time-consuming task, since the pieces should be nicely cut and matched. Therefore, automation could be a welcomed tool.


Prior work on automating collage creation extracts rectangular salient regions and assembles them in various fashions. This produces beautiful collages, however, since the extracted regions are rectangular, the variety of possible compositions is limited and many non-salient pixels are included. This approach to assemblage, while informative, does not match in spirit the way in which many artists construct collages. Artists extract the expressive regions of interest, which can be of arbitrary shape. The critical boundaries of the important information are considered significant and are thus maintained. This artistic form of collage has gained popularity also among amateurs as can be seen by the hundreds of collage groups on Flickr and hundreds of thousands users in Polyvore, an interactive web-based collage application.


We propose a method for automating collage creation, which is inspired by artistic collage work and glues mostly the purposeful cutouts. The fundamental difference between prior work and ours is that we compose a puzzle-like collage of arbitrary shaped images rather than rectangular ones. A user study shows that this creates collages that are often considered more appealing. Moreover, this allows us to generate space-efficient collages, which are useful for summarization of image data sets.


Our research makes the following contributions: a novel framework for photo-collage, a new algorithm for saliency-map computation, a Region-Of-Interest (ROI) extraction algorithm, which manages to extract non-rectangular regions that coincide with the meaningful information and an assembly algorithm that composes the ROIs in a puzzle-like manner.