Automatic Image Annotation Using CMRM with Scene Information

TELKOMNIKA, Vol.15, No.2, June 2017, pp. 693~701
ISSN: 1693-6930, accredited A by DIKTI, Decree No: 58/DIKTI/Kep/2013
DOI: 10.12928/TELKOMNIKA.v15i2.5160  693
Received January 10, 2017; Revised March 28, 2017; Accepted April 13, 2017
Automatic Image Annotation Using CMRM with Scene
Information
Julian Sahertian*
1
, Saiful Akbar
2
1
Departement of Informatics Engineering, University of Nusantara PGRI Kediri, Kampus 2,
Mojoroto Gang I, No.6, Mojoroto, Kediri, Jawa Timur, Indonesia
1,2
School of Electrical Engineering and Informatics, Bandung Institute of Technology
Jl. Ganesa No. 10, Bandung, Jawa Barat, Indonesia 40132
Corresponding author, email: juliansahertian@unpkediri.ac.id*
1
, saiful@informatika.org
2
Abstract
Searching of digital images in a disorganized image collection is a challenging problem.
One step of image searching is automatic image annotation. Automatic image annotation refers
to the process of automatically assigning relevant text keywords to any given image, reflecting
its content. In the past decade many automatic image annotation methods have been proposed
and achieved promising result. However, annotation prediction from the methods is still far from
accurate. To tackle this problem, in this paper we propose an automatic annotation method
using relevance model and scene information. CMRM is one of automatic image annotation
method based on relevance model approach. CMRM method assumes that regions in an image
can be described using a small vocabulary of blobs. Blobs are generated from segmentation,
feature extraction, and clustering. Given a training set of images with annotations, this method
predicts the probability of generating a word given the blobs in an image. To improve annotation
prediction accuracy of CMRM, in this paper we utilize scene information incorporate with
CMRM. Our proposed method is called scene-CMRM. Global image region can be represented
by features which indicate type of scene shown in the image. Thus, annotation prediction of
CMRM could be more accurate based on that scene type. Our experiments showed that, the
methods provides prediction with better precision than CMRM does, where precision represents
the percentage of words that is correctly predicted.
Keywords: automatic image annotation, CMRM, scene information
Copyright © 2017 Universitas Ahmad Dahlan. All rights reserved.
1. Introduction
In the past decade, the digital images have been increasing dramatically with the rapid
development of digital cameras and smartphones that can easily capture images. Hence, there
is a critical demand for an efficient and effective method that can help users to manage their
large volume of image. In order to organize and search images efficiently, content-based image
retrieval (CBIR) [1, 2] was proposed. CBIR approach utilizes visual features of the image such
as color, texture or shape to obtain relevant images based on user query using query-by-
example technique. However, this query technique is less popular for some users. In the other
hand, semantic search with a text-based search (eg. document search) still being user’s
preferred way to search images.
The problem when searching images using text-based search approach is an image
that we capture with digital cameras or smartphones doesn’t have textual semantic information.
To accommodate text-based search, an image requires textual annotation that defines its
semantic content information. Image annotation can be done manually or automatically. The
manual annotation is not preferable because of involving a lot of efforts. There were several
approaches to automatic image annotation e.g. classification approach [3, 4] nearest neighbour
approach [5, 6] and statistical approach called relevance model [7-10]. Compared to other
methods, relevance model was fairly robust method used in automatic image annotation.

 ISSN: 1693-6930
TELKOMNIKA Vol. 15, No. 2, June 2017 : 693 – 701
694
In this paper we propose an automatic annotation method using relevance model and
we utilize scene information to improve annotation prediction accuracy. Relevance model
assumes that image description pairs are generated by independently chosing visual features
for images and words for description. Using a training set of annotated images, joint distribution
of visual features and words is estimated. If we use contextual information of visual features,
there is a better chance of predicting correct words for description. One of the contextual
information is scene. Human vision can easily identify image scene like ‘forest’, ‘beach’, etc [11].
Using that scene information, text annotation from detail image objects can be more specifically
obtained. For example, object ‘tree’ would be more suitable annotation of scene ‘forest’, and
object ‘sand’ or ‘sea’ would be more suitable annotation of scene ‘beach’. Tariq et al. proposed
the automatic image annotation model that utilizes scene information [12]. However, they use
block-based segmentation for describing image, which is intuitively poor choice for describing
objects. Our model is inspired by the model proposed by Tariq et al., but instead of using block-
based image regions we use object-shaped regions using automatic image segmentation [13].
This paper is organized as follows. Section 2 explains the image description. Section 3
explains automatic image annotation method. Section 4 explains experiment result and
analysis. Section 5 explains conclusion and future works.
2. Image Description
Image is described from local regions using blobs representation [7]. To obtain blobs,
firstly image is divided into several regions. Then, its region visual features such as color,
texture, and shape, are extracted and clusterred. Figure 1 shows the process to obtain blobs.
Image Segmentation Clustering
Preprocessing
Features
Extraction
Blobs
Figure 1. The process of obtaining image blobs
2.1. Image Preprocessing
Image preprocessing aims to address lack of contrast between objects and noise
problems that cause poor image segmentation result. The problem of lack of contrast between
objects is solved using histogram equalization method. Histogram equalization is a method that
improves the contrast in an image, in order to stretch out the intensity range. The purpose of
histogram equalization is to obtain a wider and more uniform distribution of intensity values
histogram. The noise problem is solved by performing median filter to image. Median filter has
the same functionality as the mean filter on reducing the noise. However, different from median
filter mean filter also reduces the detail of the image which is not a noise.
2.2. Image Segmentation
Segmentation algorithm used in this paper is meanshift color segmentation [13].
Meanshift image segmentation algorithm is similar with k-means image segmentation algorithm.
Different from k-means, meanshift does not require the initiation number of clusters and only
requires one input parameter. In general, each image has a varying number of different regions,
such that for segmenting image meanshift is more suitable than k-means techniques that
require the initiated number of segment. Moreover, meanshift is a very robust technique used to
image segmentation. L*a*b* color space and Local Binary Pattern [14] image feature is used to
get better representation of image pixels. Image segmentation process using meanshift is
shown in Figure 2.

TELKOMNIKA ISSN: 1693-6930 
Automatic Image Annotation Using CMRM with Scene Information (Julian Sahertian)
695
2.3. Feature Extraction
We have implemented a set of features to describe image regions comprising color,
texture, and shape. We will describe and discuss the reasons of selecting each image features
as the followings.
2.3.1. Color Histogram
Histogram from RGB, Opponent, HSV, and Transformed Color Distribution color space
[15] is used to describe color of image regions. The histogram is quantized uniformly in 16 bins
and the resulting feature vectorper region has dimension of 48. The histogram is normalized
and equalized in order to get invariant property to image scale and illumination.
RGB
Image
Convert to
L* a* b*
L* a* b* x y
Local Binary
Pattern
Feature
Vector
Meanshift
Segmented
Image
Figure 2. Meanshift image segmentation process
2.3.2. Local Binary Pattern
Circular uniform Local Binary Pattern [14] is used to describe texture of image regions.
There are two parameters for extracting this features that are pixel number and radius. In this
reaserach we use 24 pixel numbers and 8 radius based on several experiment conducted. The
feature is represented as of histogram of unique label values of each pixel local binary pattern.
Similar to the color histogram features, the histogram is normalized in order to get it invariant to
image scaling.
2.3.3. Moments
Hu-moments [16] are used to describe the shape of image regions. Hu-moments
consist of seven invariant moments. To perform moments feature extraction, firstly we calculate
three types of moments: raw moments, central moments, and normalized central moments.
Then, we calculate seven invariant moments from normalized central moments.
3. Automatic Image Annotation Method
In this paper, in order to describe objects, we use CMRM method to automatic image
annotation [7] instead of using MBRM [9] that uses block-based regions. The following is a brief
explanation of the CMRM method we used.
3.1. CMRM Automatic Image Annotation
CMRM is a method of automatic image annotation using relevance model approach. To
annotate unlabelled image in this method we estimate joint probability between image region
and text annotation. For example, we have the training data of labelled images and test data of
unlabelled image. Firstly, each training image is segmented, and then its features such as color,
texture and shape are extracted. In order to get blobs representation, the features are
clusterred. Thus, each training image in the training dataset is represented as a set of blobs and
words. For each unlabelled test images, the image segmentation and blobs generate are also
performed. At the training stage, words probability and blobs probability are calculated for each
image in the training data T using equations (1) dan (2) [7].

 ISSN: 1693-6930
696
(1)
(2)
Where, denotes the number of times word w occurs in the caption of J, and
denotes the number of times word w occurs in the caption of all image in T. is the number
of times blob b occurs in image J, and is the number of times blob b occurs in all image
in T. is the sum of all the keywords and blobs in image J, and is the sum of all the
keywords in all image in T. and are smoothing parameters.
To annotate unlabeled test image is to estimate conditional probability
. Training data T is used to estimate the joint probability as
shown in equation (3) [5].
(3)
is considered to have a uniform distribution for all the images in T. The process for
obtaining annotation words of unlabeled images using CMRM method is shown in Figure 3.
Image
Dataset
’Words
Visual Blobs
Similarity
Test Image
Image Prior
P(J)
Image to Word
P(w|J)
Image to Image
P(b|J)
Final
Annotation
Figure 3. Image annotation using CMRM method
3.2. Scene-CMRM Automatic Image Annotation
In this subsection, we show how to utilize scene information to enhance the prediction
accuracy of CMRM. Our automatic image annotation method is called scene-CMRM. Similar to
CMRM, in order to annotate unlabeled image using scene-CMRM we estimate the joint
probability between image region and text annotation. However, this estimation is calculated
based on scene type of that image. The process of image annotation using scene-CMRM is
illustrated in Figure 4.
Image
Dataset
’Words
Visual Blobs
Similarity
Test Image
Image Prior
P(J)
Image to Word
P(w|J)
Image to Image
P(b|J)
Final
Annotation
Scene
Detection
Scene Prior
P(Ck|0)
Figure 4. Proposed image annotation method

697
Mathematical explanation of the proposed automatic image annotation system is as
follows. Assume that each image is described from n number of visual unit called blobs
. Each image also has a number of textual descriptions . In
addition we assume there are various types of scene and selection of blobs
and text annotation depends upon the type of scene at the time of generation of image pairs
and their descriptions. Thus, the annotation process of this method is to estimate the joint
probability of b and w conditional on the variable θ, where is test image scene description.
The estimation is shown in equation (4).
(4)
and from CMRM method changes as shown in (5) dan (6).
(5)
(6)
We replace in the equation (2), which is the sum of all training set is replaced, with
which is the sum of all image on a cluster scene . The final stage is sorting words in
descending order based on the probability value. Text annotations are obtained by taking n top
number of words with highest probability value or above a certain threshold.
Scene is used as prior knowledge in this proposed automatic image annotation system
obtained from scene feature extraction of image. Descriptor is used for scene detection is GIST
[11]. GIST feature is able to describe the image scene without going through the process of
objects detection in the image. GIST features describe the spatial layout of the image derived
from the global feature that is spatial envelope.
After extracting GIST feature, all images are clustered using k-means clustering
algorithm to form k clusters. For each cluster i consist of images J represented as with the
size of cluster for each scene type . Thus, for each the probability of selecting image J
conditional is shown in (7).
(7)
For each unlabeled test image, GIST featurea are also extracted from it. This feature is
variable for the unlabelled image. The probability of selecting scene considered uniform for
all scene type.
4. Experiments
This section explains the details of the dataset as well as the experiment scenarios and
evaluation measures.
4.1. Dataset
The dataset used for the experiment is image dataset from Microsoft Research in
Cambridge (MSRC) [17]. This dataset contains 591 images in 20 categories, and about 80% of
the image has more than one label annotation with an average 3 words per image. In total, the
vocabulary in this dataset contains 23 words.
In the experiments, image dataset is divided into three parts: 256 images for training, 40
images for evaluation, and 295 images for testing. The evaluation is a process of estimating
smoothing parameters. Once the parameters fixed, evaluation data and training data are
merged into new training data with 296 images or 50% of the entire dataset. Thus, at the end,
the dataset is divided into two parts: 50% for training data and 50% for testing data, as done in
[18]. Each image is manually assigned as part of training dataset or testing dataset such that
the images are proportionally distributed into the two dataset with respect to their category.

 ISSN: 1693-6930
698
4.2. Evaluation Measures
In order to measure the performance of the proposed method, we used word precision,
recall, and f-measure described as the followings.
(8)
(9)
(10)
Where is the number of word i that correctly predicted, is the number of word i that wrongly
predicted, and the number of word in test images.
5. Results and Analysis
In this section we will discuss the experimental results of our automatic image
annotation called Scene-CMRM compared to CMRM [7] as the baseline annotation method.
For both methods, the experimental results show that the increasing of word prediction
increase the precision but decrease recall. The precisions produced by scene-CMRM is slightly
better than the precision of CMRM in various number of word prediction as shown in Figure 5.
However, the recall of scene-CMRM is less than those of CMRM when the number of word
prediction is more than 3 words as shown in Figure 6. This is because the estimation of
predicted word in scene-CMRM narrowed down to specific types of scene, where the images
and words that are similar are grouped together so the percentage to predict the correct word
becomes higher.
As shown in Figure 7, we highlight that the f-measures of both methods achieve the
highest value when the number of predicted word per image is 4. As commonly knowns, the f-
measure represents the harmonics value between precision and recall. We also note that the f-
measure of scene-CMRM is higher than those of CMRM.
Figure 5. Precision of various number
predicted word
Figure 6. Recall of various number predicted
word
Figure 7. F-Measure of various number predicted word
0.0
0.1
0.2
0.3
0.4
0.5
0.6
1 2 3 4 5 6 7 8 9 10
Precision
Number of Words
cmrm
scmrm
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1 2 3 4 5 6 7 8 9 10
Recall
Number of Words
cmrm
scmrm
0.0
0.1
0.2
0.3
0.4
0.5
1 2 3 4 5 6 7 8 9 10
F-Measure
Number of Words
cmrm
scmrm

699
The detail results of per word precision and recall are shown in Figure 8 and Figure 9.
Figure 8. Per word precision
Figure 9. Per word recall
The average of per word precision, recall and f-measure of few words with high
frequency on Scene-CMRM and CMRM shown in Table 1. The examples of annotation results
from both methods are shown in Table 2.
Table 1. Performance comparison of automatic image annotation method
Method CMRM [5] Scene-CMRM
Per word Precision 0.333 0.353
Per word Recall 0.611 0.584
Per word F-Measure 0.410 0.438
Table 2. Annotation examples
Images Manual Annotation CMRM [5] Scene-CMRM
grass sky building
grass water building
sky
grass building water
sky
grass sky building
tree aeroplane
grass sky building
cat
tree building grass
sky
grass water cow
book grass water
cow
grass water cow
dog
sky road building tree
building sky road
grass
building sky road
tree
0.0
0.1
0.2
0.3
0.4
0.5
0.6
grass
sky
road
building
tree
water
body
face
cow
car
bird
sheep
flower
book
boat
bicycle
sign
chair
dog
aeroplane
mountain
cat
horse
Precision
cmrm
scmrm
0.0
0.2
0.4
0.6
0.8
1.0
1.2
grass
sky
road
building
tree
water
body
face
cow
car
bird
sheep
flower
book
boat
bicycle
sign
chair
dog
aeroplane
mountain
cat
horse
Recall
cmrm
scmrm

 ISSN: 1693-6930
700
6. Conclusion and Future Works
In this paper we proposed an alternative method to perform automatic image
annotation. In the proposed method, we modify the CMRM method by utilizing scene
information. Based on the experimental results, the proposed automatic image annotation
method called scene-CMRM provides better precision than CMRM but provides worse recall
than CMRM. This is because the estimation of predicted word in scene-CMRM narrowed down
to specific types of scene, where similar images and words are grouped together so the
percentage to predict the correct word becomes higher.
In the future, in addition to the scene information as prior knowledge we can consider
the relationship between words by utilizing lexical database like WordNet to enrich words
annotation results. Image description can be improved especially in image segmentation to get
image regions that close to ideal conditions. The use of more complex visual features like
SHIFT should be considered in further works in order to get better representation of regions.
References
[1] Kusrini K, M Dedi I, Ferry Wahyu W. Multi Features Content-Based Image Retrieval Using
Clustering and Decision Tree Algorithm. TELKOMNIKA Telecommunication Computing
Electronics and Control. 2016; 14(4): 1480-1492.
[2] Agus ZA, Rizka WS, Dimas Fanny HP, Dini AN. Region Based Image Retrieval Using
Ratio of Proportional Overlapping Object. TELKOMNIKA Telecommunication Computing
Electronics and Control. 2016; 14(4): 1608-1616.
[3] E Chang, G Kingshy, G Sychay, G Wu. CBSA: Content-Based Soft Annotation for
Multimodal Image Retrieval Using Bayes Point Machines. IEEE Trans. on CSVT. 2003;
13(1): 26–38.
[4] C Cusano, G Ciocca, R Schettini. Image annotation using SVM. Proceedings of the SPIE.
San Jose. 2003; 5304: 330-338
[5] Ameesh Makadia, Vladimir Pavlovic, Sanjiv Kumar. Baselines for Image Annotation.
International Journal of Computer Vision. 2010; 90(1): 88-105.
[6] Xirong Li, CGM Snoek, Marcel Worring. Learning Social Tag Relevance by Neighbor
Voting. IEEE Transactions on Multimedia. 2009; 11(7): 1310-1322.
[7] J Jeon, V Lavrenko, R Manmatha. Automatic Image Annotation and Retrieval Using Cross-
Media Relevance Models. Proceedings of the 26th Annual International ACM SIGIR
Conference on Research and Development in Information Retrieval. Toronto, Canada.
2003; 119-126.
[8] Victor Lavrenko, R Manmatha, Jiwoon Jeon. A Model for Learning The Semantics of
Pictures. Advances in Neural Information Processing Systems. 2003; 16: 553-560.
[9] SL Feng, R Manmatha, V Lavrenko. Multiple Bernoulli Relevance Models for Image and
Video Annotation. Proceedings of the 2004 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition. Washington DC, USA. 2004; 2: II-1002-II-1009.
[10] Hengame Deljooi, Ahmad R Eskandari. A Novel Semantic Statistical Model for Automatic
Image Annotation Using the Relationship between the Regions Based on Multi-Criteria
Decision Making. IJECE International Journal of Electrical and Computer Engineering.
2014; 4(1): 37-51.
[11] Aude Oliva, Antonio Torralba. Modeling The Shape of The Scene: A Holistic
Representation of The Spatial Envelope. International Journal of Computer Vision. 2001;
42(3): 145-175.
[12] Amara Tariq, Hassan Foroosh. Scene-Based Automatic Image Annotation. IEEE
International Conference on Image Processing (ICIP). Paris. 2014; 3047-3051.
[13] Dorin Comaniciu, Peter Meer. Mean Shift: A Robust Approach Toward Feature Space
Analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2002; 24(5):
603-619.
[14] Timo Ojala, Matti Pietikainen, Topi Maenpaa. Multiresolution Gray-scale and Rotation
Invariant Texture Classification with Local Binary Patterns. IEEE Transactions on Pattern
Analysis and Machine Intelligence. 2002; 24(7): 971-987.
[15] Koen EA Van De Sande, Theo Gevers, Cees GM Snoek. Evaluating Color Descriptors for
Object and Scene Recognition. IEEE Transactions on Pattern Analysis and Machine
Intelligence. 2010; 32(9): 1582-1596.

701
[16] Ming-Kuei Hu. Visual Pattern Recognition by Moment Invariants. IRE Transactions on
Information Theory. 1962; 8: 179-187.
[17] A Criminisi. Microsoft Research Cambridge (MSRC) Object Recognition Pixel-wise
Labeled Image Database (Version 2). 2004.
[18] Jiayu Tang. Automatic Image Annotation and Object Detection. Doctoral Thesis.
Southampton: University of Southampton, Electronics & Computer Science; 2008.

Automatic Image Annotation Using CMRM with Scene Information

More Related Content

What's hot (17)

Similar to Automatic Image Annotation Using CMRM with Scene Information (20)

More from TELKOMNIKA JOURNAL (20)

Recently uploaded (20)

Automatic Image Annotation Using CMRM with Scene Information