Procedia Computer Science 85 ( 2016 ) 954 – 961
1877-0509 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the Organizing Committee of CMS 2016 doi: 10.1016/j.procs.2016.05.287
ScienceDirect
* Corresponding author. Tel.: +91-981-597-2140;
E-mail address:[email protected]
Biomedical image indexing and retrieval descriptors: A comparative study
Gagan Deep
a*, Lakhwinder Kaur
b, Savita Gupta
caDepartment of CSE, IET Bhaddal, PTU, Ropar, India bDepartment of CE, Punjabi University, Patiala, India cDepartment of CSE, UIET, PU, Chandigarh, India
Abstract
This paper focuses on the comparison of two new proposed pattern descriptors i.e., local mesh ternary pattern (LMeTerP) and directional local ternary quantized extrema pattern (DLTerQEP) for biomedical image indexing and retrieval. The standard local binary patterns (LBP) and local ternary patterns (LTP) encode the gray scale relationship between the center pixel and its surrounding neighbors in two dimensional (2D) local region of an image whereas the former descriptor encodes the gray scale relationship among the neighbors for a given center pixel with three selected directions of mess patterns which is generated from 2D image and later descriptor encodes the spatial relation between any pair of neighbors in a local region along the given directions (i.e.,Ͳ, Ͷͷ, ͻͲ and ͳ͵ͷ) for a given center pixel in an image. The novelty of the proposed descriptors is that they use ternary patterns from images to encode more spatial structure information which lead to better retrieval. The experimental results demonstrate the superiority of the new techniques in terms of average retrieval precision (ARP) and average retrieval rate (ARR) over state-of-the-art feature extraction techniques (like LBP, LTP, LQEP, LMeP etc.) on three different types of benchmark biomedical databases.
Keywords:Medical imaging; Image retrieval; Local binary pattern (LBP); Local ternary pattern (LTP); local mesh ternary pattern; directional local ternary
quantized extrema pattern; Texture;
1. Introduction
Nowadays, texture based features play an important role as powerful discriminating visual features. They are extensively used in the image processing applications to identify the visual patterns. Some of texture feature extraction methods are proposed1,2,3,4,5,6,7,8,9. Most of biomedical images represented in gray scale are extensively textured in common. Hence, in clinical
exams, the appearance of organ/tissue/lesion in the images is due to intensity variations that laid on various texture characteristics. Thereafter, texture became very popular in biomedical image retrieval because of eminent significance of acquired texture10,11,12.
However, the computation complexity of the texture features calculated from above texture descriptors is more expansive. To address the same, the local binary pattern (LBP) is proposed13. LBP being having low computational complexity and capacity of coding minute specifications, they14 proposed some more moderations in LBP for texture classifications. Many researchers have extensively worked on LBP in various useful applications15,16,17. In the literature, many extended versions of LBP to obtain new image features for biomedical image indexing and retrieval have proposed18-24. As we know that the LBP provides the first-order directional derivative patterns but they25 examined LBP as non-directional first-order local patterns and have proposed local
derivative patterns (LDP) for face recognition. Both LBP and LDP in the literature review cannot appropriately find out the appearance distinction of particular objects in natural images due to intensity variations. To handle this, local ternary pattern (LTP) introduced for texture classification26. LTP has small pixel value variations as compared to LBP. LBP, LDP and LTP
capture the feature information based on the distribution of edges which are coded using only two directions (positive direction or negative direction). In some more researches, the local quantized patterns (LQP) for visual recognition27 and local quantized extrema patterns (LQEP) for natural and texture image retrieval have proposed28. The local mesh patterns (LMeP) have been
proposed for biomedical image retrieval23. The standard LBP method provides the connection between the referenced pixel and
its neighbors in circumference. In contrast to LBP, LQEP method provides the connection between any pair of neighbors in a local region for a given referenced pixel in an image and LMeP method provides the connection between the surrounding neighbors for a given referenced pixel in an image.
After the research papers review above, in this paper, the local patterns; LBP, LTP and LMeP motivated us to propose the LMeTerP for biomedical image indexing and retrieval. It has already been proved7,8,9 that the directional features are very
significant in various applications of image retrieval. But, the most of above literature on LBP features and its improved versions gives non-directional features. So, in this paper, the local 2D patterns; LBP, LTP and LQEP also motivated us to propose the DLTerQEP for biomedical image indexing and retrieval. Both descriptors use the ternary patterns of images to provide more spatial structure information which lead to better retrieval. The performance of the proposed descriptors is tested by conducting the experiments on three different types of benchmark biomedical databases.
2. Proposed Methods
2.1. Local mesh ternary patterns (LMeTerP)
The idea of local patterns (the LBP, the LTP, and the LMeP) has been adopted to define LMeTerP. To extend the standard LTP to LMeTerP, we generated the three local mesh pattern images at j=1, 2, 3 using convolution operations from a given image as follows: ̴݉ݐ݁ݎ̴ܽݐݐ݁ݎ݊ݏ ǡோൌ ݂ሬሬሬԦሺܿ݊ݒሺ݉݁ݏ̴݄ܮܤܲǡ ݓሼͳǡ ݆ሽǡ Ԣݏܽ݉݁ԢሻǢଶ ቀುమቁିଵ ୀଵ (1) From Eq. (1), ݂ሬሬሬԦሺݔሻ is the three value code which contains -1 or 0 or 1. This mess ternary code is converted into two binary ଶ
codes (upper LTP code and lower LTP code) at different thresholds (in this paper, 2 and -2) using the concept of LTP26. The ternary coding patterns are obtained as follows:
̴݉ݐ݁ݎ̴ܽݐݐ݁ݎ݊ݏ௨ሺͳǡ ݆ ͳሻ ൌ ݂ሬሬሬԦሺݔሻǢଶ ቀು మቁିଵ ୀଵ (2) where ݂ଶ ሬሬሬԦሺݔሻ ൌ ൜ͳǡ݂݅ሺݔ ሺݐ݄ݎ݁ݏ݄݈݀ ൌ ʹሻሻǢ Ͳǡ݂݅ሺݔ ൏ ሺݐ݄ݎ݁ݏ݄݈݀ሻሻǢ ̴݉ݐ݁ݎ̴ܽݐݐ݁ݎ݊ݏ௪ሺͳǡ ͷ ݆ሻ ൌ ݂ሬሬሬԦሺݔሻǢଶ ቀುమቁିଵ ୀଵ (3) where ݂ଶ ሬሬሬԦሺݔሻ ൌ ൜ͳǡ݂݅ሺݔ ሺݐ݄ݎ݁ݏ݄݈݀ ൌ െʹሻሻǢ Ͳǡ݂݅ሺݔ ሺݐ݄ݎ݁ݏ݄݈݀ሻሻǢ
Now, by multiplying with the binomial weights to each LTP coding, the unique LMeTerP values (decimal values) for a particular selected mess pattern (3×3) image for characterization of spatial structure of the local pattern are defined by Eq.4:
ܮܯ݁ܶ݁ݎܲఈǡൌ ̴݉ݐ݁ݎ̴ܽݐݐ݁ݎ݊ݏሺ௨ ௪ሻǡ௪Τ ିଵ
௪ୀ
ʹ௪
(4) Each LTP (upper and lower) map of mesh images of LMeTerP method is with values ranging from (0 to 2P-1). The complete image is represented by constructing a histogram using the following Eq. 5 after detecting the local pattern, PTN (LBP or LTP or
m_ter_patterns or LMeTerP). ܪெ்ଵ ሺݒሻ ൌ ͳ ݇ଵൈ ݇ଶ ݂ଶሺܮܯ݁ܶ݁ݎܲሺݍǡ ݎሻǡ ݒሻǢ మ ୀଵ భ ୀଵ ݒ א ሾͲǡ ሺʹרܲ െ ͳሻሿǢ ݂ଶሺݔǡ ݕሻ ൌ ൜ ͳǡ݂݅ሺݔ ൌ ݕሻǢ Ͳǡ݂݅ሺݔ ് ݕሻǢ (5) where k1 × k2 represents the size of an input image.
Further, we consider the uniform two operations on two LBPs which are calculated from the LTP. Further information about the uniform two patterns are as follows. For local pattern with P neighboring pixels, there are 2P (0 to 2P-1) possible values for
LBP, resulting in a feature vector of length 2P. A high computational cost is involved in extracting such a feature vector. Thus, uniform patterns17 are considered to reduce the computational cost. Thus, the distinct uniform patterns for a given query image would be P (P-1) + 2. The possible uniform patterns for P=8 can be seen17. In the above equation 5, the map values are updated
with the ranging:
ݒ א ሾͲǡ ܲሺܲ െ ͳሻ ʹሻሿǢ
Here, as we know that LTP26 is affected with the large size of feature dimensions (from 2P to 3P) (using 3-value ternary coding). In this paper using thresholds ((t) = 2, -2), it has been studied that they26 solved the above said problem by converting
the ternary pattern into two binary patterns (upper LTP and lower LTP) (see in Eq. 2 and 3).
2.2. Directional local ternary quantized extrema patterns (DLTerQEP)
The idea of local patterns (the LBP, the LTP, and the LQEP) has been adopted to define directional local ternary quantized extrema pattern (DLTerQEP). DLTerQEP describes the spatial structure of the local texture in ternary patterns using the local extrema and directional geometric structures.
In the proposed DLTerQEP for a given image, the local extrema in all directions is acquired by computing local difference between the center pixel and its neighbours by indexing of the patterns with pixel positions. The positions are indexed in a manner to write the four directional extremas operator calculations. Local directional extrema values28 (LDEV) for the local pattern neighbourhoods of an image (I) are calculated as follows:
ܮܦܧܸሺݍǡ ݎሻ ൌ ሺܫሺݍ െ ݆ǣ ݍ ݆ǡ ݎ െ ݆ǣ ݎ ݆ሻ െ ܫሺݍǡ ݎሻሻ మି ୀ భି ୀ (6) where the size of input image is k1× k2 and the values of i and j are 4, 3 respectively for getting values of directional patterns of at
least 7×7 pattern of an image.
Four directional HVDA7 geometric structure of possible directional LQP geometries is used for feature extraction. Directional
local extrema values in 0o, 45o, 90o and 135o directions are extracted in the shape of HVDA7 from LDEV values as calculated by
Eq.6. Then four directional ternary extrema codings (DTEC) are collected based on the four directions (0o, 45o, 90o and 135o) at different thresholds (in this paper, 2 and -2) using the concept of LTP26. The ternary coding patterns are obtained as follows:
ܦܶܧܥͳ൫ܫሺ݃ሻ൯หן ൌ ە ۖ ۔ ۖ ۓ ݂ሬሬሬԦ൫ܮܦܧܸሺ݃ସ ସହሻ ൈ ܮܦܧܸሺ݃ସଷሻǡ ܫሺ݃ሻ൯Ǣ݂ሬሬሬԦ൫ܮܦܧܸሺ݃ସ ସሻ ൈ ܮܦܧܸሺ݃ସଶሻǡ ܫሺ݃ሻ൯Ǣ݂ሬሬሬԦ൫ܮܦܧܸሺ݃ସ ସሻ ൈ ܮܦܧܸሺ݃ସଵሻǡ ܫሺ݃ሻ൯Ǣןൌ Ͳ ݂ସ ሬሬሬԦ൫ܮܦܧܸሺ݃ଷସሻ ൈ ܮܦܧܸሺ݃ହସሻǡ ܫሺ݃ሻ൯Ǣ݂ሬሬሬԦ൫ܮܦܧܸሺ݃ସ ଶସሻ ൈ ܮܦܧܸሺ݃ସሻǡ ܫሺ݃ሻ൯Ǣ݂ሬሬሬԦ൫ܮܦܧܸሺ݃ସ ଵସሻ ൈ ܮܦܧܸሺ݃ସሻǡ ܫሺ݃ሻ൯Ǣןൌ Ͷͷ ݂ସ ሬሬሬԦ൫ܮܦܧܸሺ݃ଷହሻ ൈ ܮܦܧܸሺ݃ହଷሻǡ ܫሺ݃ሻ൯Ǣ݂ሬሬሬԦ൫ܮܦܧܸሺ݃ସ ଶሻ ൈ ܮܦܧܸሺ݃ଶሻǡ ܫሺ݃ሻ൯Ǣ݂ሬሬሬԦ൫ܮܦܧܸሺ݃ସ ଵሻ ൈ ܮܦܧܸሺ݃ଵሻǡ ܫሺ݃ሻ൯Ǣןൌ ͻͲ ݂ସ ሬሬሬԦ൫ܮܦܧܸሺ݃ଷଷሻ ൈ ܮܦܧܸሺ݃ହହሻǡ ܫሺ݃ሻ൯Ǣ݂ሬሬሬԦ൫ܮܦܧܸሺ݃ସ ଶଶሻ ൈ ܮܦܧܸሺ݃ሻǡ ܫሺ݃ሻ൯Ǣ݂ሬሬሬԦ൫ܮܦܧܸሺ݃ସ ଵଵሻ ൈ ܮܦܧܸሺ݃ሻǡ ܫሺ݃ሻ൯Ǣןൌ ͳ͵ͷ (7) where
gij represents ith row and jth column coefficient index,
For upper LTP, DTEC1 is computed from Eq. (7) by obtaining value of ݂ሬሬሬԦସ as follows:
݂ସ
ሬሬሬሬԦሺݔǡ ݃ሻ ൌ ൜
ͳǡ݂݅ሺݔ ሺݐ݄ݎ݁ݏ݄݈݀ ൌ ʹሻሻǢ Ͳǡ݂݅ሺݔ ൏ ሺݐ݄ݎ݁ݏ݄݈݀ሻሻǢ
(8)
݂ସ
ሬሬሬԦሺݔǡ ݃ሻ ൌ ൜
ͳǡ݂݅ሺݔ ሺݐ݄ݎ݁ݏ݄݈݀ ൌ െʹሻሻǢ Ͳǡ݂݅ሺݔ ሺݐ݄ݎ݁ݏ݄݈݀ሻሻǢ
(9)
The DTEC is defined from the Eqs.7 to 9 as follows:
ܦܶܧܥ൫ܫሺ݃ሻ൯ ൌ
ܦܶܧܥͳ൫ܫሺ݃ሻ൯หǡ ܦܶܧܥͳ൫ܫሺ݃ሻ൯หସହǡ ܦܶܧܥͳ൫ܫሺ݃ሻ൯หଽǡ ܦܶܧܥͳ൫ܫሺ݃ሻ൯หଵଷହǢ
ܦܶܧܥʹ൫ܫሺ݃ሻ൯หǡ ܦܶܧܥʹ൫ܫሺ݃ሻ൯หସହǡ ܦܶܧܥʹ൫ܫሺ݃ሻ൯หଽǡ ܦܶܧܥʹ൫ܫሺ݃ሻ൯หଵଷହ
൩
(10) As above DTEC coding is converted into two binary codes (upper LTP code and lower LTP code) as that in the LTP26. So in practice, DLTerQEP allows concatenation of four directional extremas for P=12-bit (w=0...11) binary coding string generation (3-bit from each given directional extrema) in each binary pattern of LTP.
Now, by multiplying with the binomial weights to each DTEC LTP coding, the unique DLTerQEP values (decimal values) for a particular given pattern (7×7) for characterization of spatial structure of the local pattern are defined by Eq.11.
ܦܮܶ݁ݎܳܧܲఈǡൌ ܦܶܧܥሺ௨ ௪ሻǡ௪Τ ିଵ
௪ୀ
ʹ௪
(11) For whole image, each DTEC LTP (upper and lower) map from above is with values ranging from 0 to 4095 (0 to 2P-1), so complete DTEC LTP map is built with values ranging from 0 to 8191(0 to ((2(2P
))-1). The complete image is represented by constructing a histogram using the following Eqs. 12 and 13 after detecting the local pattern, PTN (LBP or LTP or DTEC or DLTerQEP). ܪ்ொாଵ ሺݒሻ ൌ ͳ ݇ଵൈ ݇ଶ ݂ଶ൫ܦܮܶ݁ݎܳܧܲ௨ሺݍǡ ݎሻǡ ݒ൯Ǣ మ ୀଵ భ ୀଵ ݒ א ሾͲǡ ͶͲͻͷሿǢ ݂ଶሺݔǡ ݕሻ ൌ ൜ ͳǡ݂݅ሺݔ ൌ ݕሻǢ Ͳǡ݂݅ሺݔ ് ݕሻǢ (12) ܪ்ொாଶ ሺݒሻ ൌ ଵ భൈమσ σ ݂ଶሺܦܮܶ݁ݎܳܧܲ௪ሺݍǡ ݎሻǡ ݒሻǢ ݒ א ሾͲǡ ͶͲͻͷሿǢ మ ୀଵ భ ୀଵ (13)
Parameter specifications of Eq.13 are same as in Eq.12,
ܪ்ொாሺݒሻ ൌ ሾܪ்ொாଵ ሺݒሻǡ ܪ்ொாଶ ሺݒሻሿ (14)
The proposed DLTerQEP is disparate from the familiar method LBP. The DLTerQEP extracts the spatial relation between any pair of neighbors in a local region along the given directions, while LBP extracts relation between the center pixel and its neighbors. DLTerQEP take out the directional edge information based on local extrema that differ it from the existing LBP. Therefore, DLTerQEP captures more spatial information as compared with LBP. DLTerQEP also provides a significant increase in discriminative power by allowing larger local pattern neighborhoods.
Fig. 1 illustrates the responses obtained by applying the LBP, the LTP, the LMeTerP and the DLTerQEP on a reference face image. Face image is chosen as it provides the results which are visibly comprehensible to differentiate the effectiveness of these approaches. It is monitored that the feature map of DLTerQEP (upper and lower LTPs) operator is able to capture more directional edge information as compared to feature map of LBP, LTP and LMeTerP for texture extraction.
^ĂŵƉůĞŝŵĂŐĞ
>WĨĞĂƚƵƌĞŵĂƉ >DĞdĞƌW ĨĞĂƚƵƌĞŵĂƉ >dĞƌY W Ͳ ƵƉƉĞƌ
>dWĨĞĂƚƵƌĞŵĂƉ
>dĞƌY W Ͳ ůŽǁĞƌ >dWĨĞĂƚƵƌĞŵĂƉ Fig. 1. Response of proposed method on a reference face
3. Experimental results
The performance of the proposed algorithm is analyzed for biomedical image retrieval by conducting experiments on three
different medical databases. In all the experiments, each image in the database is chosen as the query image. For each query, the system collects n database images ܺ ൌ ൫ݔଵǡݔଶǡǥǡݔ൯ǡ with the shortest image matching distance. If ݔǢi=1,2,...,n associate with the
same category of the query image, we can say that the system has correctly matched the desired. The results yielded by the proposed methods are examined in the following subsections.
3.1. Experiment on LIDC-IDRI-CT data set
Experiments are conducted on the images collected from Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI), which is public lung image database of CT scans29. The database is separated into 84 cases, each containing around 100-400 Digital Imaging and Communication (DICOM) images and an XML data file containing the physicians annotations. Database contains 143 nodules having size range of 3 to 30mm (to be manually segmented by radiologists). For the experiment, 12 patient cases consisting of 75 nodules (26 benign and 49 malignant) and 229 slices have been selected.
Fig.2. Performance comparison of the proposed methods (LMeTerP and DLTerQEP) with other existing methods by passing different query images (1-10) in
terms of ARP on LIDC-IDRI-CT database.
The Fig. 2 shows the retrieval performance for top ten matches of the proposed methods (LMeTerP and DLTerQEP) and some other existing methods (LBPu2, WLD, LQP, LMeP, LQEP) in terms of ARP by passing different query images (1-10) on LIDC-IDRI-CT database. The Fig. 2 shows that LMeTerP achieves better performance over the DLTerQEP and other existing methods LBPu2, WLD, LQP, LMeP and LQEP on most of the cases.
3.2. Experiment on VIA/I-ELCAP-CT data set
To perform the evaluation of different computer-aided detection systems, computer tomography (CT) dataset is designed by
Vision and image analysis (VIA) group and international early lung cancer action program (I-ELCAP) (VIA/I-ELCAP CT lung image dataset is available online30). Experiments are performed on CT scans of about 10 scans having 100 images each of resolution 512×512.
Table I shows the performance of the proposed methods and other existing methods in terms of ARP (%) on VIA/I-ELCAP-CT database. Fig. 3 illustrates the retrieval performance of the proposed methods (LMeTerP and DLTerQEP) and other existing methods (LBP/LBPu2, INTH, GLCM1, GLCM2, LMeP, LTP, LDP, LTCoP and LQEP) in terms of ARR. From Table I and Fig. 3, it is clear that the proposed method (DLTerQEP) outperforms the LMeTerP and other existing methods in terms of ARP and ARR on VIA/I-ELCAP-CT database.
3.3. Experiment on OASIS-MRI data set
In the experimentation, the open access series of imaging studies (OASIS) database is utilized. The open access series of imaging studies (OASIS) is a series of magnetic resonance imaging (MRI) dataset. (Dataset is available online for use in medical research field31). A cross-sectional collection of 421 patients aged between 18 and 96 years is used for the experimentation. Furthermore, based on the shape of ventricular in the images, four categories (124, 102, 89, and 106 images) from 421 images are grouped to evaluate the retrieval of images.
0 0,2 0,4 0,6 0,8 1 1,2 0 2 4 6 8 10 12 ARP(%)
Query image data
LBPu2 LMeTerP LMeP WLD LQP LQEP DLTerQEP
Table I: Performance of the proposed methods and other existing methods in terms of ARP (%) on VIA/I-ELCAP-CT database. Method ARP(%) 10 20 30 40 50 60 70 80 90 100 INTH 60.76 51.325 45.8833 42.2475 39.242 36.715 34.4786 32.5438 30.8633 29.402 GLCM1 63.37 53.11 46.9567 42.9475 39.716 37.1733 34.9586 32.8938 31.1067 29.627 GLCM2 65.07 55.33 49.62 45.405 42.152 39.3483 36.9429 34.8125 32.9733 31.383 LBPu2 79.21 73.77 70.26 66.9025 64.056 61.3467 58.9314 56.7175 54.31 51.915 LTP 71.59 65.29 61.44 58.17 55.82 54.12 52.09 50.03 48.44 47.53 LTCoP 86.95 78.55 73.55 69.19 65.47 62.44 60.02 57.5 54.97 52.47 LDP 86.12 77.83 72.57 67.93 63.21 60.01 57.55 54.68 52.02 48.95 LMeP 83.28 76.5 71.97 68.06 64.83 62.05 59.5643 57.2575 54.9356 52.695 LMeTerP 88.19 80.69 75.39 71.189 67.49 64.21 61.33 58.6 56 53.34 LQEP 82.94 76.09 71.22 67.29 64.17 61.36 58.98 56.77 54.55 52.41 DLTerQEP 87.58 79.88 74.6 70.36 66.91 63.73 61.05 58.61 56.41 54.07
Fig. 3. Comparison of (LMeTerP and DLTerQEP) and other existing methods in terms of ARR on VIA/I-ELCAP-CT database.
Table II summarizes the retrieval results of (LMeTerP and DLTerQEP) and other existing methods on OASIS database in terms of ARP. Fig. 4 shows the graphs depicting the retrieval performance of the proposed method and other existing methods on different values of (P, R) in terms of ARP as function of number of top matches. From Fig. 4 and Table II, it is evident that the proposed method (DLTerQEP) outperforms the LMeTerP and the many of other existing methods on OASIS-MRI image database for biomedical image retrieval.
4 9 14 19 24 29 34 39 44 49 54 59 10 20 30 40 50 60 70 80 90 100 ARR (%)
Number of top matches considered
INTH GLCM1 GLCM2 LBPu2 LTP LTCoP LDP LMeP LMeTerP LQEP DLTerQEP
Fig. 4. Comparison of proposed methods and other existing methods on different values of (P, R) as function of number of top matches in terms of ARP on
OASIS-MRI database.
Table II: Comparison of various techniques showing group wise performance in terms of precision on OASIS-MRI database Method Precision (%) (n=10)
4. Conclusions
In this paper, new proposed descriptors for biomedical image indexing and retrieval are compared. To test the robustness and effectiveness of the proposed algorithms, results are taken on three different types of benchmark medical databases. The detailed investigations of the results show that the proposed method (DLTerQEP) outperforms in terms of ARP and ARR as compared to LBP, LTP, LMeTerP and other state-of-the-art methods on LIDC-IDRI-CT, OASIS-MRI and VIA/I-ELCAP-CT databases.
References
1. Haralick, R., Shanmugan, K., Dinstain, I., Textural features for image classification. IEEE Transactions on Systems, Man and Cybernetics 1973; 3(6): pp.610-622.
2. Choi, H., Baraniuk, R., Multiscale image segmentation using wavelet-domain hidden markov models. IEEE Trans. on Image Processing 2001; 10(9): pp.1309-1321.
3. Chen, J., Shan, S., He, C., Zhao, G., Pietikäinen, M., Chen, X., Gao, W., WLD: A Robust local image descriptor. IEEE transactions on pattern analysis and
machine intelligence 2009; 32(9): pp.1705-1720.
4. Zhang, L., Zhou, Z., & Li, H., Binary Gabor pattern: An efficient and robust descriptor for texture classification. 19th IEEE International Conference on
Image Processing (ICIP) 2012; pp.81-84.
5. Nanni, L., Lumini, A., & Brahnam, S., Local binary patterns variants as texture descriptors for medical image analysis. Artificial Intelligence in Medicine 2010; 49(2): pp.117-125.
6. Do, M., N., Vetterli, M., Wavelet-based texture retrieval using generalized Gaussian density and Kullback-leibler distance. IEEE Trans. on Image
Processing 2002; 11(2): pp.146-158.
7. Kokare, M., Biswas, P., K., Chatterji, B., N., Texture image retrieval using new rotated complex wavelet filters. IEEE Trans Syst Man Cybernet 2005; 33(6): pp.1168-1178. 35 40 45 50 55 60 65 70 75 80 85 90 95 100 1 2 3 4 5 6 7 8 9 10 ARP(%)
Number of top matches considered
LBPu2_8_1 LBPu2_16_2 LMeTerP CS_LBP BLK_LBP DBWP LTP DLEP LDP LMeP_8_1 LMePu2_8_1 LQEP DLTerQEP
Group1 Group2 Group3 Group4 Total
LBPu2_8_1 51.77 32.54 33.82 49.06 42.63 LBPu2_16_2 52.58 38.43 31.68 51.13 44.37 LMeTerP 54.03 38.04 35.73 50.09 44.47 CS_LBP 44.7 40.1 31.17 48.27 41.06 BLK_LBP 48.13 41.22 35.16 50.01 43.63 DBWP 52.74 37.74 34.38 60 47.05 LTP 56.53 36.27 34.97 50 45.17 DLEP 50.51 36.28 37.01 68.21 48 LDP 46.29 36.37 36.82 45.56 41.8 LMeP 52.82 36.56 36.08 51.5 44.96 LQEP 50.97 36.37 39.1 35.57 40.51 DLTerQEP 51.61 38.63 36.97 65.57 48.19
8. Kokare, M., Biswas, P., K., Chatterji, B., N., Rotation-invariant texture image retrieval using rotated complex wavelet filters. IEEE TransSyst Man
Cybernet 2006; 36(6): pp.1273-1282.
9. Kokare, M., Biswas, P., K., Chatterji, B., Texture image retrieval using rotated wavelet filters. J Pattern Recog. Lett 2007; 28: pp.1240-49.
10. Felipe, J., C., Traina, A., J., M., Traina, C., J., Retrieval by content of medical images using texture for tissue identification. In:Proceedings of the 16th IEEE Symposium on Computer-Based Medical Systems, New York, USA, 2003; pp.175-180.
11. Jhanwara, N., Chaudhuri, S., Seetharamanc, G., Zavidovique, B., Content based image retrieval using motif co-occurrence matrix. Image and Vision
Computing 2004; 22(14): pp.1211-1220.
12. Scott, G., Shyu, C., R., Knowledge-driven multi dimensional indexing structure for biomedical media database retrieval. IEEE Trans. Inf. Technol. Biomed. 2007; 11 (3): pp.320-331.
13. Ojala, T., Pietikainen, M., Harwood, D., A comparative study of texture measures with classification based on feature distributions. Pattern Recognition 1996; 29 (1): pp.51-59.
14. Ojala, T., Pietikainen, M., and Maeenpaa, T., Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE
Transactions on Pattern Analysis and Machine Intelligence 2002; 24(7): pp.971-987.
15. Nanni, L., and Lumini, A., Local binary patterns for a hybrid fingerprint matcher. Pattern Recognition 2008; 41(11): pp.3461-3466.
16. Heikkila, M., Pietikainen, M., Schmid, C., Description of interest regions with local binary patterns. Pattern Recognition 2009; 42: pp.425-436.
17. Guo, Z., Zhang, L., and Zhang, D., Rotation invariant texture classification using LBP variance with global matching. Pattern recognition 2010; 43(3): pp.706-716.
18. Murala, S., Maheshwari, R., P., and Balasubramanian, R., Directional binary wavelet patterns for biomedical image indexing and retrieval. J. Med. Syst. 2012; 36 (5): pp.2865-2879.
19. Murala, S., Maheshwari, R., P., and Balasubramanian, R., Directional local extrema patterns: a new descriptor for content based image retrieval. Int. J.
Multimedia Inf. Retr. 2012b; 1 (3): pp.191-203.
20. Murala, S., Maheshwari, R., P., and Balasubramanian, R., Local maximum edge binary patterns: a new descriptor for image retrieval and object tracking.
Signal Processing 2012c; 92(6): pp.1467-1479.
21. Murala, S., Maheshwari, R., P., and Balasubramanian, R., Local tetra patterns: new feature descriptor for content based image retrieval. IEEE Trans. Image
Processing 2012d; 21(5): pp.2874±2886.
22. Murala, S., and Jonathan, W., Q., Local ternary co-occurrence patterns: a new feature descriptor for MRI and CT image retrieval. Neurocomputing 2013; 119 (7): pp.399-412.
23. Murala, S., and Jonathan, W., Q., Local mesh patterns versus local binary patterns: biomedical image indexing and retrieval. IEEE Journal of Biomedical and Health Informatics 2014; 18 (3): pp.929-938.
24. Murala, S., and Jonathan, W., Q., MRI and CT image indexing using local mesh peak valley edge patterns. Signal processing: image communication 2014b; 29(3): pp.400-409.
25. Zhang, B., Gao, Y., Zhao, S., Liu, J., Local derivative pattern versus local binary pattern: Face recognition with higher-order local pattern descriptor. IEEE
Trans. Image Processing 2010; 19 (2): pp.533-54.
26. Tan, X., Triggs, B., Enhanced local texture feature sets for face recognition under difficult lighting conditions. IEEE Trans. Image Process 2010; 19(6): pp.1635-1650.
27. Hussain S., U., Triggs, B., Visual recognition using local quantized patterns. European conference on computer vision (ECCV 2012), Part-II, LNCS 7573, Italy 2012; pp.716-729.
28. Rao, L., K., Rao, D., V., Local quantized extrema patterns for content based natural and texture image retrieval. Human Centric Computing and
Information Sciences, DOI: 10.1186/s13673-015-0044-z, 2015; 5(26).
29. NEMA-CT image database, (2012) [Online]. Available: ftp://medical.nema.org/medical/Dicom/Multiframe/.
30. VIA/I-ELCAP CT Lung Image Dataset, Available [online] <http://www.via.cornell.edu/-databases/lungdb.html>.
31. Marcus, D., S., Wang, T., H., Parker, J., Csernansky, J., G., Morris, J., C., Open access series of imaging studies (OASIS): Cross sectional MRI data in young, middle aged, nondemented, and demented older adults. J. Cogn. Neurosci. 2007; 19 (9): pp.1498-1507.
