Multichannel Decoded Local Binary Patterns for Content Based Image Retrieval IEEE Transactions on Image Processing, 2016 Shiv Ram Dubey, Satish Kumar Singh and Rajat Kumar Singh Indian Institute of Information Technology, Allahabad
Method
TABLE I Truth Table of Adder and Decoder map with 3 input channels ๐ณ๐ฉ๐ท๐๐ ๐, ๐ ๐ณ๐ฉ๐ท๐๐ ๐, ๐ ๐ณ๐ฉ๐ท๐๐ ๐, ๐ ๐๐๐ด๐ (๐, ๐) ๐๐
๐ด๐ (๐, ๐)
The framework of feature description of color image using adder and decoder based multichannel LBP is shown in Fig. 1. Color Image Red (R) Channel
Green (G) Channel
Blue (B) Channel
LBP1
LBP2
LBP3
0 0 0 0 1 1 1 1
Decoder
Adder maLBP1
ยฐ ยฐ ยฐ
maLBP4
mdLBP1
ยฐ ยฐ ยฐ
mdLBP8
maLBP1 Histogram
ยฐ ยฐ ยฐ
maLBP4 Histogram
mdLBP1 Histogram
ยฐ ยฐ ยฐ
mdLBP8 Histogram
maLBP Feature Vector
mdLBP Feature Vector
0 0
0 1 1 2 1 2 2 3
1
a
0
1
0 0
0
0 0
a 0
1 0
1
1 0
1
0
LBP 2
0 0
a
1 0
LBP 3
0 1
a
0
0
0
0
0 1
MIT-VisTex database 100 LBP cLBP mscLBP mCENTRIST maLBP mdLBP
95 90 85 80 75 65 1
95 90 LBP cLBP mscLBP mCENTRIST maLBP mdLBP
85 80 75
2
3
4
5
6
7
8
Number of Retrieved Images
9
10
70 1
2
3
4
5
6
7
8
9
10
Number of Retrieved Images
Fig.4. The performance comparison of proposed maLBP and mdLBP descriptor with existing approaches such as LBP, cLBP, mscLBP, and mCENTRIST descriptors over Corel-1k and MIT-VisTex databases.
0
0
0 0
1 0
0
๐
a
0 1
๐
0 0
a
0
0
0
0
0 0
a 0
0 1
0
0
0
0
0 0
pixel
0
0
0
f (c) Weighting function
72
161
20
2
maLBP 1
maLBP 2
maLBP 3
maLBP 4
72
128
32
0
0 0
mdLBPn4
mdLBP 1
mdLBP 2
mdLBP 3
mdLBP 4
0
a
0
4
0
a
0
0
0 0
0
0
a
0 0
0
1
16
4
2
1
0
0
1
0
mdLBPn5
mdLBPn6
mdLBPn7
mdLBPn8
(f) Eight output decoder LBPs
๐ผ๐ก๐โ1 (๐ฅ, ๐ฆ)
0
0 0
a
1
0
0 0
mdLBPn3
0
๐ผ๐ก๐ (๐ฅ, ๐ฆ)
0 0
mdLBPn2
6
DM
1 2
0
a
0
0
0
๐ผ๐ก๐+1 (๐ฅ, ๐ฆ)
0
2
AM
8
(e) Multichannel adder based local binary pattern for each output channels 0
mdLBPn1 0
๐ผ๐ก1 (๐ฅ, ๐ฆ)
1
7
128
a
16
1
maLBPn4
0 1
0
4
0
maLBPn3
0 0
3
a
5
32
0
a
0
0 1
maLBPn2
1
64 1
0 0
a
1
0
a
2
2
(b) Adder map and Decoder map
(d) Four output adder LBPs
a
Corel-1k database 100
70
0 1
0
(a) Three input LBPs for three channels
maLBPn1
of
0 1 2 3 4 5 6 7
0 1
1
1
1
๐ผ๐ก2 (๐ฅ, ๐ฆ)
Fig.2. The local neighbors ๐ผ๐ก๐ (๐ฅ, ๐ฆ) ๐ผ๐ก ๐ฅ, ๐ฆ in ๐ก ๐กโ channel for ๐ โ [1, ๐] and ๐ก โ [1,3].
1
LBP 1
1
โ ๐ผ๐ก ๐ฅ, ๐ฆ
a 0
0
๐ผ๐ก๐ (๐ฅ, ๐ฆ)
0 0
1
๐ผ๐ก3 (๐ฅ, ๐ฆ)
2๐ ๐
0 1 0 1 0 1 0 1
Image retrieval experiments are performed to test the performance of proposed descriptors in terms of average retrieval precision (ARP). Precision is the percentage of correct number of retrieved images out of total number of retrieved images. Results are compared with basic LBP [1] and other color based local descriptors such as cLBP [2], mscLBP [3], and mCENTRIST [4]. Here, results are presented over natural Corel-1k [5] and textural MIT-VisTex [6] databases in Fig.4.
The computation of ๐๐๐ฟ๐ต๐๐ก 1 for โ๐ก1 โ [1,4] and ๐๐๐ฟ๐ต๐๐ก 2 for โ๐ก2 โ [1,8] from input ๐ฟ๐ต๐๐ก๐ ๐ฅ, ๐ฆ using an example of three LBP patterns is illustrated in Fig. 3 for ๐ = 8.
Fig.1. The flowchart of computation of multichannel adder based local binary pattern feature vector (i.e. maLBP) and multichannel decoder based local binary pattern feature vector (i.e. mdLBP) of an image from its Red (R), Green (G) and Blue (B) channels.
๐ผ๐ก๐ โ1 (๐ฅ, ๐ฆ)
0 0 1 1 0 0 1 1
Performance Evaluation
ARP (%)
๏ Local binary pattern (LBP) [1] is widely adopted for simplicity and efficient image feature description. ๏ To describe the color images, it is required to combine the LBPs from each channel of the image. ๏ We introduced adder and decoder based two schemas for the combination of the LBPs from more than one channel. ๏ The introduced descriptors significantly improve the retrieval performance and outperform the other multichannel based approaches.
The position of local neighbours of any pixel of the image is depicted in Fig.2. A local binary pattern ๐ฟ๐ต๐๐ก (๐ฅ, ๐ฆ) for a pixel (๐ฅ, ๐ฆ) in ๐ก๐กโ channel is generated as follows, ๐ ๐ ๐ฟ๐ต๐๐ก ๐ฅ, ๐ฆ = ๐ โ๐ก โ 1,3 (1) ๐=1 ๐ฟ๐ต๐๐ก ๐ฅ, ๐ฆ ร ๐ , where, 1, ๐ผ๐ก๐ ๐ฅ, ๐ฆ โฅ ๐ผ๐ก ๐ฅ, ๐ฆ ๐ฟ๐ต๐๐ก๐ ๐ฅ, ๐ฆ = (2) 0, ๐๐กโ๐๐๐ค๐๐ ๐ and ๐ ๐ is a weighting function defined by the following equation, ๐ ๐ = (2)(๐โ1) , โ๐ โ [1, ๐] (3) ๐ The truth map of adder map (i.e. ๐๐๐ (๐ฅ, ๐ฆ)) and decoder map (i.e. ๐๐๐ ๐ (๐ฅ, ๐ฆ)) are shown in Table 1.
ARP (%)
Introduction
mdLBP 5
mdLBP 6
mdLBP 7
mdLBP 8
(g) Multichannel decoder based local binary pattern for each output channels
Fig.3. An illustration of the computation of the adder/decoder based local binary pattern maps, adder/decoder based local binary pattern bits, and adder/decoder local binary pattern decimal values from three 8-bit input LBPs. Green and Red circles represent 0 and 1 respectively.
Fig.5. Top 10 retrieved images (columns) using LBP (1st row), cLBP (2nd row), mscLBP (3rd row), mCENTRIST (4th row), maLBP (5th row) and mdLBP (6th row) descriptors from Corel-1k database. Images in the 1st column are query images as well as the top most similar images.
References [1]T. Ojala, M. Pietikainen and T. Maenpaa, โMultiresolution gray-scale and rotation invariant texture classification with local binary patterns,โ IEEE TPAMI, 24(7): 971-987, 2002. [2]J.Y. Choi, K.N. Plataniotis and Y.M. Ro, โUsing colour local binary pattern features for face recognition,โ 17th IEEE ICIP, 2010. [3]C. Zhu, C.E. Bichot and L. Chen, โMulti-scale Color Local Binary Patterns for Visual Object Classes Recognitionโ IEEE ICPR, 2010. [4]Y. Xiao, J. Wu and J. Yuan, โmCENTRIST: A Multi-Channel Feature Generation Mechanism for Scene Categorization,โ IEEE TIP, 23(2): 823-836, 2014. [5]Corel Photo Collection Color Image Database taken from: http://wang.ist.psu.edu/docs/realted/. [6]MIT Vision and Modeling Group, Cambridge, โVision texture databaseโ taken from: http://vismod.media.mit.edu/pub/.