Static Images Project
Our research directions in CBIR
1. Detecting of color patterns library according to the human visual perception
Color features play a large role in content-based image retrieval since it is a very meaningful feature for human visual perception. And it is very important for every content-based image retrieval system to judge the similarity between colors in the same way as humans do. We address the issue of how humans perceive and measure similarity within the domain of colour patterns. To detect optimal color patterns that correspond to human perception we performed a subjective experiment. (H, S, V) color space were factorized in several ways and experiment shows the best solution. We propose also a novel method for color representation based on our color patterns library.
2. Establishing a correspondence between low-level features and semantics of fixed images
The performance of content-based image retrieval systems is still judged to be unsatisfactory. Many researches see the main reason for that in so-called “semantic gap” between low-level features and high-level content of the image. Bridging this semantic gap is known as a difficult task. We propose a novel way to establish a correspondence between low-level color-features and high-level content of images. We consider a database of natural images where no additional semantic information about images is available. To build a mapping table, basis color features were built as a first step. These basis color features describes a regularity of color distribution for groups of similar images. A training set of open-scene images was defined. The training set was clustered based on color data. Basis color features are defined as a color features mean of the same cluster’s images. As a second step, a list of primitive words (which we called lexical basis features) was selected from a lexicon of the Russian language. Then lexical basis features were used to describe a semantic content of one cluster’s images. Putting together basis color features and basis lexical features for each cluster we set up a correspondence between semantic and color content of natural fixed images.
3. Nearest-neighbor search problem in the context of CBIR
Recently, a lot of indexing methods were developed for multidimensional spaces and for image feature vector spaces in particular. We propose an implementation of a vantage point tree method. The goal is to define the parameters and algorithms for the nearest-neighbor search problem in the context of the content-based image retrieval task. We also provide some experiments to determine the dependences between index efficiency and algorithms' parameters.
4. Adaptive fusion of color and texture features depending on the query
It is a common way in CBIR to combine various candidate image features to obtain a better result set. Simple solution to combine independent evidence is to use a weighted average. We detected query-images’ classes with the same optimal weights for queries in the same class. Then we detect common features for the images from the same class. As a result we provide a mapping between query feature patterns and optimal weights for combining texture and color evidences.
5. Fusion of image features
We propose a novel method for fusion the result sets obtained based of different image features. A novel fusion-function is used for merging ranked lists with weights. First experiments shows that our method outperforms well known CombSUM, CombMAX, CombMNZ in merging homogeneous evidences. Experiments with heterogeneous evidences will follow later.
6. Image retrieval based on shape features
Many approaches exists utilizing shape representation and comparison, e.g. the methods based on Fourier descriptors, Grid-based methods, turning angle method and others. But there is no objective comparison between all these methods. Our experiments are to show which of existing methods are the best for different image training sets.
Vassilieva N., Novikov B. A Similarity Retrieval Algorithm for Natural Images. Proc. of the Baltic DB&IS'2004, Riga, Latvia, June 2004.
N. Vassilieva, B. Novikov. Establishing a correspondence between low–level features and semantics of fixed images. Proc. of the Seventh National Russian Research Conference RCDL’2005, Yaroslavl, Russia, October 2005.
I. Markov. VP-tree: Content-Based Image Indexing. Proc. of the Fours Spring Colloquium for Young Researchers in Databases and Information Sytems, SYRCoDIS’2007, Moscou, Russia, June 2007.
I. Markov, N. Vassilieva, A. Yaremchuk. Image retrieval. Optimal weights for color and texture fusion based on query object. Accepted to the Ninth Russian National Research Conference Pereslavl, Russia October 15-18, 2007.
- Natalia Vassilieva
- Maria Teplykh
- Maria Davydova
- Daria Bratchikova
- Ilya Markov (http://ilya.markov.googlepages.com/home)
- Alexander Dolnik
- Dmitri Shubakov