Professor of Artificial Intelligence, Vice-Rector for Research and Innovation
CBIR SOFTWARE · RESEARCH ARCHIVE
img(Rummager).
An interactive research platform for content-based image retrieval, visual descriptor experimentation and retrieval evaluation.
OVERVIEW
A laboratory for visual search.
img(Rummager) brings together original and established visual descriptors within a single experimental application. It can search with a query image using XML index files or work directly with an image folder, extracting comparison features at retrieval time.
The platform also supports hybrid retrieval that combines keywords with visual similarity. Evaluation facilities include Normalized Modified Retrieval Rank (NMRR), Mean Normalized Retrieval Order, average precision and export in trec_eval format.
01 · SEARCH
Query by example
Retrieve visually similar images from an index or a local folder.
02 · DESCRIBE
Visual features
Experiment with colour, texture, shape, local and composite descriptors.
03 · COMBINE
Hybrid retrieval
Combine keyword information with content-based visual similarity.
04 · EVALUATE
Retrieval metrics
Measure ranked results and export standard evaluation files.
WINDOWS · .NET FRAMEWORK 3.5 · GNU GPL
Archived downloads.
img(Rummager) is a portable C# application originally designed for Windows XP or later with .NET Framework 3.5. It does not require installation. Its XML indexes can be stored locally or on a network location.
SOFTWARE ARCHIVE
Download img(Rummager).
These packages are retained for historical and research use. Compatibility with current Windows environments is not guaranteed.
Fixed CCDs late-fusion and MPEG-7 late-fusion bugs.
Fixed colour-histogram and autocorrelogram issues, with thanks to Oge Marques.
Fixed local-folder search.
Added SMART weighting and normalisation methods for visual-word search.
Introduced the alpha version of Colour Visual Words (CoViWo).
Added custom and dynamically sized codebooks.
Added new index-file creation methods.
Stored SURF and CoViWo features in binary files for faster retrieval.
Reduced XML file size and improved batch-mode performance.
Added control over the number of results exported to TREC files.
EXPERIMENTAL WORKBENCH
The visual feature laboratory.
The Laboratory tab collects colour, texture, transform, shape, local-feature, clustering, filtering and fusion components used in image-retrieval research.
Colour and texture components +
Colour elements
RGB, YIQ, LAB, HSV, HSL and HMMD colour spaces, annular luminosity, DCT and binary Haar transforms on histograms, Hartley transforms, MPEG-7 SCD, CLD and DCD, fuzzy HSV and LAB linking, autocorrelograms, spatial colour layout and related methods.
Texture elements
Co-occurrence matrices, Laws’ texture measures, Wang’s wavelet, MPEG-7 Edge Histogram Descriptor, Tamura histograms and related techniques.
Descriptors, transforms and fusion methods +
Additional visual components
Wavelet and Fourier transforms, Hu and Zernike moments, more than 30 filters, Coordinate Logic Filters, clustering methods, face detection, CEDD, FCTH and JCD composite descriptors, colour reduction, brightness and texture directionality descriptors, SURF, visual words and shape features.
Fusion and supporting tools
Z-score, Borda variants, inverse rank position, CombSUM, fuzzy combination and other fusion strategies, together with real-time colour correction, edge extraction, filter evaluation, automatic descriptor selection, Hough transforms, fuzzy colour clustering and texture-based image tiling.
RETRIEVAL MODES
Three ways to search.
The application supports indexed retrieval, direct folder search and an experimental hybrid Flickr workflow.
01 · Search with an XML index file +
This is the recommended workflow. XML input files store descriptors for collections of images. Supported representations include CEDD, FCTH, JCD, MPEG-7 EHD, SCD and CLD, autocorrelograms, RGB histograms, Tamura directionality, fusion techniques, SURF, Visual Words and CoViWo.
Create a searchable file through Manage Index Files: select the descriptor, define the output folder and filename, and choose the folders containing the images.
The MPEG-7 code was adapted from the GPL implementation available through LIRE, and img(Rummager) was designed to reproduce its descriptor results.
02 · Search directly in a folder +
Direct folder search provides custom and preset modes. Custom search allows the researcher to select the characteristics used during retrieval, while preset search offers nine application-defined descriptors. After retrieval, NMRR can be calculated using selected relevant images or a ground-truth file.
03 · Experimental Flickr hybrid search +
The original prototype connected to a project server, downloaded available keywords and combined them with a visual query to retrieve images from Flickr. This historical integration may no longer function because it depends on external services and legacy interfaces.
RELATED PROTOTYPES
Experiments built around the platform.
The software package also included demonstration tools, classifiers and small applications that explored alternative uses of visual descriptors.
img(Paint.Anaktisi) +
A sketch-based retrieval experiment using the MPEG-7 Colour Layout Descriptor. Users could search by drawing a coarse colour layout, while the compact descriptor enabled efficient matching. The application used the same XML indexes as img(Rummager).
img(Finder) +
An early beta prototype that connected to Facebook, generated a Joint Composite Descriptor index from accessible images and allowed a user to test whether a query image appeared in that collection. The original external integration is now historical.
img(Encryption) +
A visual-content encryption experiment based on cellular automata and a Coordinate Logic Filter XOR property. A same-sized key image and one-time-pad method were combined with randomisation, while semi-blind source separation reconstructed a lossless representation of the original grayscale or colour image.
LEGACY MATERIAL
Video tutorials.
The original tutorials were published as SWF files. Modern browsers usually do not play this format directly, but the files remain available as part of the archive.
IMG(RUMMAGER): An Interactive Content Based Image Retrieval System
S. A. Chatzichristofis, Y. S. Boutalis and M. Lux. 2nd International Conference on Similarity Search and Applications, IEEE Computer Society, pp. 151-153, 29-30 August 2009, Prague, Czech Republic.
Libraries, components and acknowledgements +
The project used Emgu CV, OpenCV, JChampion, AForge.NET components by Andrew Kirillov, MPEG-7 implementations influenced by the Caliph and Emir project by Mathias Lux, the Aerosquare icon pack, Asim Goheer’s EXIF extractor library and Christ Kennedy’s rotating picture tray.
img(Rummager) was released under the GNU GPL. A commercial GUI component prevented publication of the complete interface source. Historical source-code enquiries can be sent to savvash@gmail.com.
