Pathology Image Retrieval by Block LBP Based pLSA Model with Low-Rank and Sparse Matrix Decomposition - Advances in Image and Graphics Technologies - page 333

To evaluate the perform
BoW (LPB-BoW), block-L
Caicedo et al. [6] and Kow
select 300 images of each
time, we calculate the mea
returned images through co
performances of these five m

mance of our method, we compare it with block-LBP-ba
LBP-based pLSA (LBP- pLSA), along with approaches
wal et al [7]. Note that we perform 20 times to random
category for training and the remaining for test. For e
n Average Precision (mAP) these five methods for top
osine distance based similarity measure. Table 1 shows
methods.

ased
s of
mly
each
p 20
the

Table 1. Performa ance comparison at the top 20 returns of five methods

Algorithm Performance

Kowal et al. [7] 53.1±1.61

C Caicedo et al. [6] 70.2±1.06

LPB-BoW 74.4±1.17

LBP-pLSA 76.3±0.41

Our method 78.6±0.49

Fig. 5 5. Precision-recall curves of five methods

As can be seen from Tab
al. [6] and Kowal et al [7],
due to the fact that block-L
texture structure, and in
tic characterization ability o
tation of BoW, LBP-pLSA
method is more excellent t

ble 1, compared with the methods proposed by Caicedo
, LPB-BoW has superior retrieval performance. It may
LBP features can effectively describe the spatial property
n the other hand, it may benefit from the sem
of BoW. Particularly, as pLSA model overcomes the li
A and our method are better. It should be noted that
than LBP-pLSA, since it can discover the correlated

o et
y be
y of
man-
imi-
our
and