Information Technology Reference
In-Depth Information
ϭ
Ϭ͘ϵ
Ϭ͘ϴ
ĐĐƵƌĂĐLJŽŶůĂƐƐŝĨŝĞĚ
EŽƚůĂƐƐŝĨŝĞĚWĞƌĐ͘
ŽƌƌĞĐƚůLJůĂƐƐŝĨŝĞĚ>ŽƐƚƐWĞƌĐ͘
Ϭ͘ϳ
Ϭ͘ϲ
Ϭ͘ϱ
Ϭ͘ϰ
Ϭ͘ϯ
Ϭ͘Ϯ
Ϭ͘ϭ
Ϭ
Ϭ Ϭ͘ϭ Ϭ͘Ϯ Ϭ͘ϯ Ϭ͘ϰ Ϭ͘ϱ Ϭ͘ϲ Ϭ͘ϳ Ϭ͘ϴ Ϭ͘ϵ ϭ
/ŵĂŐĞĐůĂƐƐŝĨŝĐĂƚŝŽŶĐŽŶĨŝĚĞŶĐĞƚŚƌĞƐŚŽůĚ
ɋɋɋɋ
Fig. 4.
Results obtained by the image similarity based classifier for similarity
s
h,and
using SIFT,
for various classification confidence thresholds (
ν
img
)
performs better than the other, we only report the result obtained for the
and
version of
each measures and for the best
k
.
The first observation is that the
Local Feature Based Image Classifier
(
Φ
m
)per-
forms significantly better then any
Image Similarity Based Classifier
. In particular
Φ
m
performs better then
s
h,and
, even if no geometric consistency checks are performed by
Φ
m
while matches in
s
h,and
are filtered making use of the Hough transform.
Even if in this paper we did not consider the computational cost of classification,
we can make some simple observations. In fact, it is worth saying that the local feature
based classifier is less critical from this point of view. First, because closest neighbors
of local features in the test image are searched once for all in the
Tr
and not every time
for each image of
Tr
. Second, because it is possible to leverage on global spatial index
for all the features in
Tr
, to support efficient
k
nearest neighbors searching. In fact,
the similarity function between two local features is the Euclidean distance, which is a
metric. Thus, it could be efficiently indexed by using a metric data structures [18,15,4].
Regarding the local features used and the computational cost, we underline that the
number of local features detected by the SIFT extractor is twice that detected by SURF.
Thus, on one hand SIFT has better performance while on the other hand SURF is more
efficient.
8
Conclusions
In this paper we addressed the problem of image content recognition using local features
and kNN based classification techniques. We defined 20 similarity functions and com-
pared their performance on a image content landmarks recognition task. We found that
a two-way comparison of two images based on fuzzy
and
allows better performance
than the standard approach that compares a query image with the ones in a training set.
Moreover, we showed that the similarity functions relying on matching of local features
that makes use of geometric constrains perform slightly better than the others.
Finally, we defined a novel kNN classifier that first assigns a label to each local
feature of an image and then labels the whole image by considering the labels and the
confidences assigned to its local features.
