It is widely acknowledged that good performances of content-based image retrieval systems can be attained by adopting relevance feedback mechanisms. One of the main difficulties in exploiting relevance information is the availability of few relevant images, as users typically label a few dozen of images, the majority of them often being non-relevant to user’s needs. In order to boost the learning capabilities of relevance feedback techniques, this paper proposes the creation of points in the feature space which can be considered as representation of relevant images. The new points are generated taking into account not only the available relevant points in the feature space, but also the relative positions of non-relevant ones. This approach has been tested on a relevance feedback technique, based on the Nearest-Neighbor classification paradigm. Reported experiments show the effectiveness of the proposed technique relatively to precision and recall.
Design of robust classifiers for adversarial environments - Systems, Man, and...
Wiamis2010 poster
1. K-NEAREST NEIGHBORS DIRECTED SYNTHETIC IMAGES INJECTION
Luca Piras, Giorgio Giacinto
Dept. of Electrical and Electronic Engineering - University of Cagliari, Italy
luca.piras@diee.unica.it - giacinto@diee.unica.it
K-Nearest Neighbors Directed Pattern Injection
Relevance feedback We propose to address the problem of small number of training
Relevance feedback techniques are employed to capture the patterns by artificially increasing the number of relevant patterns in
subjectivity of retrieval results and to refine the search. The user is asked the feature space. The basic idea underlying K-Nearest Neighbors
to label the images as being relevant or not, and the similarity measure Directed Pattern Injection consists in adding some artificial patterns
is modified accordingly: the images in the database are ranked in generated in the feature space by taking into account the K
terms of the distances from the nearest relevant and non-relevant relevant images nearest to a reference point in the feature space.
images.
I ! NN nr I () 1 K
$ # X % Qk ( )
relevanceNN I = () Pi = X + ! "
()
I ! NN r I + I ! NN nr I () K k =1 k
The choice of all parameter is far from being a trivial task; the
NN r
( I ) ( resp., NN ( I ))nr
nearest relevant (resp., non-relevant) image to
image I.
effectiveness of the method depends from:
X Reference relevant image K n° Nearest Neighbors ! k ! N (0,1)
- Query
Digital Compute the new
Library query using the - mean vector of all the relevant images (Mr)
user's hints Non-relevant Relevant - each relevant images (Ar)
- a point computed according to the BQS New artificial
query movement technique (BQS) pattern
Compute the similarity between
the query image and the
images in the database
Example-Image ! Scaling factor
chosen by a user to
1
perform a search in a !=
Digital Library
The user marks the relevant and non-relevant
("1
2
+ " 22 )
Bound of linear
images for the relevance feedback 1
!= combination
("1 + "2 )
After several
iterations 1 Bound of nearest
!= relevant image
("1 + "2 )
2
Bound of farthest
relevant image
The system outputs the results
Qk k-th relevant image nearest to X
N where n = 2, …, 5
Number of pattern Pi : =n
R+ P
Relevant images found at
the previous steps
Lack of relevant images
One of the most severe problems in exploiting relevance feedback in
image retrieval is the small number of images that the user considers as
being “relevant” compared to the number of non-relevant images
especially during the first iterations. Experimental results
The Caltech-256 dataset obtained from the California Institute of
I) Even in the case of very “cooperative” users, it is not feasible to Technology repository has been used. It consists of 30607 images
display more than a few dozens of images to label. subdivided into 257 semantic classes and the number of images per
class ranges from 80 to 827.
II) If the database at hand is very large, then the number of images
that are relevant to the query can easily be very small compared to 500 images have been randomly extracted from all of the 257 classes,
the size of the database. and used as query. The top twenty nearest neighbors of each query
III) In a high dimensional feature space images with different are returned. 9 relevance feedback iterations are performed.
semantic content can lie near each other.
The Tamura representation has been used (18-dimensional feature vector).
1
Our proposal F=
1 1
Increasing the number of “relevant” patterns used to train the system: +
2 ! prec 2 ! recall
I) Creating new random artificial patterns by exploiting nearest
neighbor information.
II) Constraining these patterns in a region of the feature space
containing relevant images.
Pattern Recognition and Applications Group
P
R A Group http://prag.diee.unica.it/pra/eng/home
WIAMIS 2010