- Planing of the work for the testing phase of the thesis.
- Research for existing codes in the web.
- Implementation of some cost function methods in matlab
- Building of a matlab UI for the cost functions implemented until the moment
- Incorportion of seam carving in the matlab UI
- Metting with supervisor and co-supervisor, at 6/Mar, to show what was done until the date, and to receive feedback
- Inclusion of another cost function to the matlab UI
- Inclusion of DTW algorithm to the matlab UI (still need a few enhancements)
- Sending emails to some article authors in order to know if they can provide their code
- Corrections in DTW code
- Meeting with supervisor and co-supervisor, at 18/Mar, to show what was done until now and to talk about some problems along the dissertation
- Inclusion of seams selection tool in matlab UI
- Meeting with supervisor, at 25/Mar, to talk about some doubts in the DTW algorithm
- Corrections of the DTW, discussed in the 25/Mar metting, made
- Original DTW code already working
- Meeting with supervisor, at 10/Abr, to show the DTW algorithm, and discussion of ways to inovate it
- Begining of inovating the DTW algorithm with mask
- Search of ways to use metrics in the resizing case
- Meeting with supervisor and co-supervisor, at 17/Abr, to talk about the metric and some inovation viability.
- Initiation of the implementation of objective metrics in the work
- Meeting with supervisor, at 24/Abr, to find ways to correct a problem in the DTW (when not having stable paths, how to chose the paths to be taken)
- Attempt to implement object removal in the algorithm (not working)
- Enhancement of the metrics implementation
- Meeting with supervisor, at 30/Abr, to talk about the INESC demo on the 6/May and to find ways to make the DTW algorithm faster
- Participation in INESC demo day at 6/May
- Adition of new code to the DTW algorithm
- Final enhancements of the metrics code and inovations to the DTW algorithm
- Structuring of the monography
- Writing of the some metodology in the monography
- Continuation of the monography metodology writing
- Writing of the results in the monography
- Meeting with supervisor, at 22/May, to talk about the monography structuration and the final algorithm
- Continuation of the monography results writing
- changes in the monography metodology writing
- Continuation of the monography metodology writing
- Meeting with supervisor and co-supervisor to a brief analysis of the monography writing
- Reewriting of a Results section
- Writing of abstract, conclusion and future work
- Meeting with supervisor for analysis of the monography
- Final changes in the monography
- Delivery of the CD with the monography
With the appearing of several display devices, such as mobile phones and computer monitor, a necessity to resize images emerged. As the typical resizing methods, like the simple scaling and cropping, bring undesired essential content elimination, details softening, stretching and squeezing, the retargeting methods appeared. These methods methodology seeks to change an image size while preserving the important content.
The main goal of this research was the implementation of an accurate image retargeting algorithm that reduces and enlarges images without causing the feeling of a damaged image, maintaining the core elements of the image and eliminating others with least importance. The first step for this work, was the development of an intuitive Matlab application, essentially for the testing process. This process passed through two stages: finding the best cost function, that allows to change the image size while preserving its main content, and improvement of the Stable Paths algorithm accompanied with the enlargement tool adding.
In order to make the fairest possible subjective selection and taking into account the limited existing time, a database, from Rubinstein et al. [50], was used as a starting point, aiming for an image resizing with the same reducing/enlarging factors used by these authors. The cost function was selected, having as parameters the complete retarget of the image to the new specified measure, with the minimal artifacts and distortions, using the human perceiving system. With this, our methodology was applied to several images for both reducing and enlarging, followed by the application of the metrics measures to all, including on some of the results provided by Rubinstein et al. [50], that were also addressed in the Literature Review.
Then, after an analysis of the obtained metrics data, it was concluded that our method, Seam Carving and the Nonhomogeneous Warping methods gave the closest result values to the ideals of translation, scaling and rotation, between an object in the retargeted and original images. It was also observed, that, by little, our method was the one closest to the ideal values of object features preservation.
Despite the relatively upbeat results, the method still lacks precision, causing some images retargets areas to lack structure, especially in aleatory lines that appear in backgrounds, where less important image contents are, and in more geometric structures. That appends, because similarly to seam carving, these images content is laid out in a way that prevents seams from bypassing those structures.