Changing an Image Size without Information Loss (No. 1461)
Lead Researcher: Prof. Michal Irani
Bidirectional Similarity offers a new approach to summarization of visual data (images and video) based on optimization of well defined similarity measure.
Common visual summarization methods (mainly scaling and cropping) suffer from significant deficiencies related to image quality and loss of important data. Many attempts have been made to overcome these problems, however, success was very limited and neither has become commercially applicable.
Using an optimization problem approach and state-of-the-art algorithms, our method provides superior summarization of visual data as well as a measure to determine similarity, which together provides a basis for a wide range of applications in image and video processing.
The technology can be utilized in any application where an image size is changed or were similarity of images is important. Sample applications include:
- Image processing software (as an added-on feature)
- Resizing software
- Creation of Thumbnails
- Adjustment of images to different screen sizes (TV-cellular etc.)
- Optimization of space-time patches in video processing
- Image montages
- Automatic image & video cropping
- Images synthesis, photo reshuffling and many more
While Bidirectional Similarity summarization will not replace existing technologies in all applications, it enjoys significant advantages that will offer better results in many of them. Among its advantages, the Bidirectional Similarity summarization:
- Provides better resolution and in many cases reduces distortion compared to scaling
- Reduces (or avoids) loss of important data compared to cropping
- Allows importance-based summarization even when important information is widespread and hard to define
- Uses quantitative objective similarity measure
- Offers a generic tool for different image processing applications (synthesis, montage, reshuffling etc.)
Bidirectional Similarity Summarization is a patent-pending image and video processing method, which maximizes “completeness” and “coherence” between images and videos, using a measure for quantifying how “good” a visual summary is.
The algorithm uses and iterative process, gradually reducing the image size, while keeping all source patches in the target image, without introducing visual artifacts that are not in the input data. Using a Similarity Index, the algorithm identifies redundant information and compromise the “less important” data while generating the required target image or video.
The Similarity Index, which stands in the heart of the Bidirectional Similarity summarization algorithm, can be utilized by its own, as an objective function within other optimization processes, as well as in comparing the quality of visual summaries generated by different methods