https://doi.org/10.31449/inf.v45i7.3715 Informatica 45 (2021) 97–105 97 
 
Effective Image Representation using Double Colour Histogram for 
Content-Based Image Retrieval 
Ezekiel Mensah Martey, Hang Lei and Xiaoyu Li 
School of Information & Software Engineering, 
University of Electronic Science and Technology of China, China 
E-mail: martey003@std.uestc.edu.cn, hlei@uestc.edu.cn, xiaoyu33521@163.com 
 
Obed Appiah 
Department of Computer Science and Informatics,  
University of Energy and Natural Resource, Ghana 
Email: obed.appiah@uenr.edu.gh 
Keywords: stacked colour histogram, conventional colour histogram, texton, image representation, content-based image 
retrieval 
Received: September 1, 2021 
Image representation is critical to the successful realisation of Content-Based Image Retrieval (CBIR) 
systems. The choice of features to represent the image affects retrieval performance. Nowadays, image 
databases are heterogeneous and different feature types can be used for appropriate descriptions. This 
paper proposes an image representation for CBIR that combines Stacked Colour Histogram (SCH) and 
Conventional Colour Histogram (CCH) to improve image retrieval precision. This presented technique is 
designed to capture the colour and texture information of the image. The colour properties of an image 
are represented by CCH and that of texture by SCH. The weighted similarity measure is used to estimate 
the proportion of similarity values in the retrieval task. The novel descriptor has been widely tested on 
four standard image datasets, namely Batik, Coil100, Corel10K and Outext. Batik, Coil100 and Outext 
are used to assess texture discrimination. Corel10K is used to assess the discrimination of heterogeneous 
images. Experimental results and comparisons with SCH, CMTH, MTH, TCM, CTM and NRFUCTM 
demonstrate that the proposed descriptor has superior retrieval performance. 
Povzetek: V članku je podana učinkovita predstavitev slik z barvnim diagramom z namenom iskanja 
vsebine na slikah. 
 
1 Introduction 
Image representation is crucial in achieving high precision 
retrieval results. The image is represented based on low-
level features that are colour, texture and shape. Among 
them, colour is the most commonly used to retrieve 
images. Colour feature has extensive applications in 
numerous fields such as internet image search, image 
analysis, remote sensing, medical image detection, and 
video surveillance. Conventional Colour Histogram 
(CCH)[1], Colour Coherent Vector (CCV) [2], Colour 
Moment (CM)[3] and Colour Correlogram (CC)[4] are 
standard colour feature extraction techniques that are used 
for indexing images in CBIR applications. CCH is much 
more flexible to use among the colour feature extraction 
techniques because of its simplicity and low 
computational time complexity.  However, the CCH lacks 
information about the spatial layout of colour distributions 
in an image, making it challenging to represent texture or 
shape characters of the image [5]. Notwithstanding, CCH 
and other histogram image representation techniques offer 
a great advantage of effectively handling scaling, rotation, 
reflection and translation transformation that an image 
may go through [6]. This advantage has resulted in the use 
of CCH and its variants still active in the domain of CBIR. 
To extract colours and their neighbourhood 
information, image texture feature extractions are largely 
employed.  Most CBIR techniques will use this approach 
because it tends to deal with the weaknesses of the CCH 
and hence performs better mostly when the application is 
not so much interested in the distribution of pixels but the 
location of pixels, especially in the medical image 
retrieval applications [7]. Texture defines a spatial layout 
of pixels in an image area, and many researchers have 
studied it [8]. Edge Histogram Descriptor (EHD)[9], Local 
Binary Patterns-based descriptors (LBP) [10,11], Gray-
Level Co-occurrence Matrix (GLCM) [12], Gabor filter 
[13,14], Curvelet transform[15] and Hidden Markov 
random field [16] are some texture descriptors used for 
CBIR. Texture descriptors do not always yield optimal 
results because they turn to be invariant to image 
transformations. 
The popular term texton proposed in [17] has played 
a significant role in texture analysis. The purpose of 
textons is to use basic primitives to describe complex 

98 Informatica 45 (2021) 97–105 E. M. Martey et al.  
 
structures. Textons have been applied in a variety of ways, 
such as seen in Multi-Texton Histogram (MTH) [18], 
Texton Co-occurrence Matrix (TCM) [19], Complete 
Texton Matrix (CTM) [20] and Complete Multi Texton 
Histogram [21] with generally good results.  However, the 
rigid nature of the textons makes it undesirable at some 
point [22].  This is generally because when the orientation 
of an image changes, the texton feature extractions also 
turn to produce different feature maps or vectors for the 
same image. The SCH approach proposed in [23] 
employed recurrent blurring to extract inherent textural 
information, thus solving the rigid nature of the textons. 
A single feature extracted from colour or texture 
cannot be sufficient enough to provide the best retrieval 
results. Therefore, hybrid feature extraction techniques 
have become an adequate alternative in dealing with such 
problems associated with individual feature extraction 
issues. As the images dataset increases in size and variety, 
a single feature extraction technique for indexing images 
for images search is not enough. One approach to increase 
the CBIR performance is to utilise fusion methods. Liu 
and Yang [19] fused colour feature information and pixel 
orientation into an image representation vector while [24–
28] integrated the colour and texture information in a 
single descriptor using textons.  
In this paper, we advanced the scheme Stacked Colour 
Scheme (SCH) [23] by integrating it with Conventional 
Colour Histogram (CCH) as these features are robust and 
compact for heterogeneous image representation and 
invariant to transformations (rotation and translation) [29, 
30].  
To this end, we propose a novel feature representation 
technique that joins two features into a single vector for 
indexing images. Therefore, our proposed technique 
provides; 1) a double colour histogram for an image 
representation vector that effectively deals with colour and 
heterogeneity of texture images. 2) compact information 
about colour and texture features with high precision. To 
test the effectiveness of the proposed technique, we 
compared with well-known image descriptors for CBIR. 
2 Proposed methodology 
The proposed technique is based on combining two colour 
histograms, CCH and SCH. Detailed information on the 
integration of CCH and SCH features are discussed in the 
subsections. 
2.1 Conventional Colour Histogram  
CCH [1] is the most utilised method to represent the colour 
distribution of an image. The colour space is required to 
compute the CCH. The CCH is constructed in RGB colour 
spaces. “The colour bins are represented in histograms, 
and the total number of bins in RGB histograms can be 
determined based on RGB planes” [31- 33]. 
Algorithm 1: Conventional Colour Histogram 
Algorithm 
1. I  Read an image 
2. q(I) //quantized colour 
3. cH(I)  Extract histogram of the image I  
4. cH(I) //Normalize colour histogram  
5. Store cH as an Index for the read image.  
2.2 Stacked Colour Histogram (SCH) 
SCH is the bin-by-bin summation of several histograms 
generated from the recurrent transformation of an image 
[23].  Mean filter is used to transform images iteratively. 
The filtering, which employs a window size of n   × n, is 
performed N times [34]. In each instance of the filtering 
process, the output of the previous filtering is used as input 
to that of the current filtering processes. Histograms are 
generated from filtered images and later added together to 
estimate the SCH. The filtering process uses 
neighbourhood information and therefore estimated 
histogram contains information about the nature of pixels' 
spatial distribution in images.  
Algorithm 2: Stacked Colour Histogram Algorithm 
1. I  Read an image 
2. sH (J) // Histogram with J bins.  Initialize with Zeros. 
3. count  N  // N is the number of times the filtering of 
image is to be done. 
4. while (count > 1) 
5.    H(J)  Extract histogram of image I 
6.    sH(binIndex) = sH(binIndex) + H(binIndex) 
7.    while (binIndex  >  1) 
8.       sH(binIndex) = sH(binIndex) + H(binIndex) 
9.       binIndex = binIndex − 1  
10.   I  Mean Filter (I) and assign to I. //Perform 
mean filtering of the image in variable I 
11. count = count – 1 
12. Store sH as Index for the read image.  
2.3 Proposed joint histogram based on 
SCH and CH 
The diagrammatic representation of the proposed 
approach based on CCH and SCH features is shown in 
Figure 1. Concise descriptions of the key steps are as 
follows: 
Image indexing 
1. Read the image (I) 
2. Quantised image into 64 colours 
3. Compute CCH of (I) 
4. Compute SCH of (I)  
5. Combine CCH and SCH into a single vector and 
store as an image's Index.  The index vector 
should have two parts, LowerIndex which 
contains the CCH, and UpperIndex, which stores 
SCH 
Image retrieval 
1. Read the image (I) Query Image 
2. Quantised image into 64 colours 
3. Compute CCH of (I) 
4. Compute SCH of (I)  
5. Combine CCH and SCH into a single vector 

Effective Image Representation using Double Colour Histograms for ... Informatica 45 (2021) 97–105 99 
 
6. Estimate the similarity of LowerIndex of Query 
image with the LowerIndex of Target image 
using the Euclidean distance and store value as 
LI 
7. Estimate the similarity of UpperIndex of Query 
image with the UpperIndex of Target image 
using the Euclidean distance and store value as 
U.I. 
Using weighted values (𝛼 , 𝛽 ), estimate the final similarity 
value by equation 1 
        Similarity = 𝛼 ∗ 𝐿𝐼 + 𝛽 ∗ UI                           (1) 
The final similarity is used to retrieve images that are 
closer in the distance to the query image. 
𝛼 𝑎𝑛𝑑 𝛽 𝑎𝑟𝑒 𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑢𝑠𝑖𝑛𝑔 𝑒𝑞𝑢𝑎𝑡𝑖𝑜𝑛 2 ∶  
                           𝛼 + 𝛽 = 1                                             (2) 
such that             0 ≤ 𝛼 ≤ 1 
0 ≤ 𝛽 ≤ 1 
The Euclidean distance is presented in equation 3: 
d
E
  =√∑ (p
i
-q
i
)
2 n
i=1
                                                     (3)                        
where:   
               𝑑 𝐸   = distance metric 
                𝑃 𝑖   = target image 
               𝑞 𝑖   = query image 
3 Experimental evaluation 
For evaluation of our experiment, we used four popular 
datasets used in [23]—Batik, Corel10K, Outext and 
Coil100. The Batik, Coil100 and Outext datasets are used 
to assess the proposed descriptor in differentiating texture 
images. The Corel10K dataset is used to assess the 
heterogeneous image recognition capabilities of the 
proposed technique.  
Below is a concise description of datasets. 
❖ Batik   [35] dataset contains 300 images in total. The 
image database is categorized into 50 classes, and 
each class contains six images. The images are of 
128x128 pixels sizes in JPEG format. 
❖ Coil100 [36] dataset contains 7200 colour images 
categorised into 100 classes. There are 72 images of 
each object in different poses in each class.  
❖ Outext [37] dataset contains 11484 total images. The 
image database is categorised into 29 classes.  
❖ Corel10K [38] comprises 10000 images in total. 
There are 100 classes, and each class contains 100 
images. The images are sized 128 × 192 pixels in 
JPEG format.  
Samples images from each dataset are illustrated in 
Figure 2. 
The initial experiment evaluated window sizes and 
recurrent transformations on accuracy using different 
classifiers and different dataset features. The proposed 
method implemented here used a window size of 7x7 and 
11x11 with recurrent transformations of 5, 10, 15, 20 and 
25 to extract features for this experiment. 
 
 
 
Figure 2: Samples images from each dataset (a) Batik 
(b) Outext (c) Coil100 and (d) Corel10K. 
 
Figure 1: Block diagram of the proposed technique. 

100 Informatica 45 (2021) 97–105 E. M. Martey et al.  
 
Machine learning approach for evaluation  
K-fold cross-validation was performed to avert the over-
fitting of the data during the process of training. The 
different classifiers with K-fold cross-validation (k   = 5) 
were used to evaluate the performance on multiple and 
different features of datasets. Using Principal Component 
Analysis (PCA), we reduced the feature sizes to speed up 
computation, setting the threshold to 95%. 
Five different classifiers, namely Discriminant 
Analysis Classifier (DAC), Support Vector Machine 
(SVM), Naive Bayes (NB), Decision Tree (DT) and KNN 
(k-Nearest Neighbours), were used to classify images 
from each dataset. Classification accuracy metric is used 
to measure the performance of classifiers. Equation (4) 
indicates mathematical expression for classification 
accuracy. 
    𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 
𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑙𝑦 𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑑 𝑖𝑚𝑎𝑔𝑒𝑠 𝑡𝑜𝑡𝑎𝑙 𝑖𝑚𝑎𝑔𝑒𝑠                      (4) 
3.1 Image retrieval   
The performance of the proposed descriptor is evaluated 
in the classic CBIR task. Precision and Recall metrics are 
used to evaluate the performance of the proposed 
descriptor in the CBIR task. The precision determines the 
number of correctly retrieved images. Equations (5) and 
(6) present the Precision and Recall metrics formulae, 
respectively. 
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑟𝑒𝑡𝑟 𝑖𝑒𝑣𝑒𝑑 𝑟𝑒𝑙𝑒𝑣𝑎𝑛𝑡 𝑖𝑚𝑎𝑔𝑒𝑠 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑟𝑒𝑡𝑟𝑖𝑒𝑣𝑒𝑑 𝑖𝑚𝑎𝑔𝑒𝑠          (5) 
Recall=
number of retrieved relevant images
Number of relevant images in the dataset
   6) 
We first present the results of the machine learning 
approach used for the experiment. 
Table 1 presents the mean accuracy values of the 
KNN classifier with different values of k on three datasets. 
The mean accuracies when k = 1, 98%, 38.28%, 79.48% 
were recorded for Coil100, Corel10K, and Outext 
datasets, respectively. When k = 3, mean accuracy 
percentages of 96.24%, 37.26%, 78.08% were recorded 
for Coil100, Corel10K, and Outext datasets, respectively. 
The value k = 5 produced accuracy values of 95.35%, 
39.82%, 76.64% for Coil100, Corel10K, and Outext 
datasets, respectively. Finally, using k = 9 recorded 
95.15%, 40.34%, 77.94% for Coil100, Corel10K, and 
Outext datasets, respectively. The mean accuracies give a 
good indication of the classification efficiency of the 
proposed scheme. 
The results obtained for the KNN classifier were 
compared with features in [29]. The accuracy percentages 
of our proposed descriptor show improved performance in 
representing heterogeneous texture images. Also, our 
method performs very well for texture images. Table 2 
compares the performance of our method with other 
approaches for different values of k (KNN's parameter) 
To experimentally validate if our proposed method 
keep up with other classifiers as implemented in our 
earlier work [29], we reconduct evaluation using Decision 
Tree (DT), Discriminant Analysis Classifier (DAC), 
Support Vector Machine (SVM), and Naive Bayes (NB).  
The best accuracy values achieved by the DT classifier on 
all datasets are shown in table 3. The Coil100, Corel10K, 
Outext and batik datasets recorded best accuracy values of 
87.5%, 20.5 %, 66.3% and 51.6% respectively. Mean 
accuracies of 84.78%, 23.3%, 65.38% and 39.18% were 
k-Value Coil100 Corel10K    Outext 
k=1 0.98 0.3828 0.7948 
k=3 0.9624 0.3726 0.7808 
k=5 0.9535 0.3982 0.7664 
k=9 0.9515 0.4034 0.7794 
Table 1: accuracy values for KNN classifier with different 
values of k on three datasets. 
  Outext Corel10K 
 TCM 72.49% ± 2.8% 16.71% ± 0.1% 
 MTH 51.92% ± 2.3% 23.68% ± 0.1% 
 CTM 72.31% ± 2.8% 21.22% ± 0.2% 
 NRFU 36.72% ± 1.7% 5.38% ± 0.0% 
 CMTH 78.40% ± 3.0% 31.76%± 0.1% 
 
SCH 77.63%± 3.0% 31.80%± 0.2% 
k = 1 
Proposed    
Technique  
79.48%± 2.0% 38.28%± 0.1% 
 TCM 76.42% ± 3.0% 16.29 % ± 0.1% 
 MTH 52.30% ± 2.4% 22.68 % ± 0.2% 
 CTM 70.18% ± 2.8% 20.02 % ± 0.2% 
 NRFU 36.89% ± 1.8% 5.26 % ± 0.0% 
 CMTH 76.42% ± 3.0% 30.89 %± 0.2% 
 SCH 77.33%± 3.0% 31.35%± 0.2% 
k = 3 
Proposed 
Technique 
78.08%± 2.0% 37.26%± 0.1% 
 TCM 68.41 % ± 2.7% 17.51 % ± 0.1% 
 MTH 52.68 % ± 2.5% 25.16 % ± 0.2% 
 CTM 68.93 % ± 2.9% 20.97 % ± 0.2% 
 NRFU 39.49 % ± 2.1% 6.00 % ± 0.08% 
 CMTH 74.83 % ± 3.0% 33.16 % ± 0.2% 
    
 
SCH 76.04%± 3.0% 34.03% ± 0.2% 
k = 5 
Proposed 
Technique 
78.08% ± 2.0% 39.82% ± 0.1% 
 TCM 72.49% ± 2.8% 16.71% ± 0.1% 
 MTH 51.92% ± 2.3% 23.68% ± 0.1% 
 CTM 72.31% ± 2.8% 21.22% ± 0.2% 
 NRFU 36.72% ± 1.7% 5.38% ± 0.0% 
 CMTH 78.40% ± 3.0% 31.76% ± 0.1% 
 SCH 77.19%± 3.0% 31.93% ± 0.2% 
k = 9 
Proposed 
Technique 
77.94% ± 2.0% 40.34% ± 0.1% 
Table 2: Results of proposed technique compared with 
other approaches for different values of k (KNN’s 
parameter). 
 

Effective Image Representation using Double Colour Histograms for ... Informatica 45 (2021) 97–105 101 
 
recorded for Coil100, Corel10K, and Outext datasets 
respectively. 
Table 4 demonstrates the outcomes of the SVM 
classifier on the datasets. The best accuracy result of 
97.4% is obtained for Coil100. The transformations of 20 
produced the best classification result of 50.3 % for 
Corel10K. Accuracy percentages of 78.9% and 91.7% are 
produced as best for the Outext and batik datasets, 
respectively. Mean accuracy percentages of 96.4%, 
49.57%, 77.58% and 86.98% were obtained for Coil100, 
Corel10K, Outext and batik datasets, respectively. 
Table 5 illustrates the outcome of the NB classifier 
analysed on the four datasets. The classifier produced the 
highest accuracy of 93.7% for Coil100. 41 % is generated 
as the highest for Corel10K. 63.9% is produced as the best 
accuracy percentage for the Outext dataset. Mean 
accuracies of 92.7%, 40.20%, and 63.12% were recorded 
for Coil100, Corel10K and Outext datasets, respectively. 
The results obtained by the DAC classifier are shown 
in table 6. The accuracy values of 93.2%, 44.2 %, and 
70.2% for the respective Coil100, Corel10K, and Outext 
datasets demonstrate the classification efficiency of the 
proposed scheme. Mean accuracies of 92.2%, 43.91%, 
and 69.22% were recorded for Coil100, Corel10K, and 
Outext datasets respectively. 
Table 7 presents the results compared for four different 
classifiers for well-established SCH and texton based 
feature extraction techniques. The texton-based features 
and SCH lost performances in representing heterogeneous 
images. Our proposed method outperforms the SCH and 
textons. (i.e., SCH yields 32.72% with Corel10K. 
However, our proposed technique outperforms with 
43.91% for DAC). 
We present the results of the image retrieval 
performance 
The proposed technique is primarily proposed for 
indexing images for CBIR. We evaluated the proposed 
descriptor in a classic CBIR task. The experiment initially 
evaluated the effect of several window sizes and the 
number of recurrent transformations on Precision and 
Recall on the identified dataset. The 7x7 window size with 
Window Size 
− Recurrent 
Number Coil100 Corel10K Outext 
7x7 - 5 0.834 0.23 0.781 
7x7 - 10 0.845 0.232 0.771 
7x7 - 15 0.84 0.231 0.762 
7x7 - 20 0.845 0.231 0.722 
7x7 - 25 0.875 0.241 0.783 
Table 3: accuracy results for DT classifier on four datasets. 
Window Size − 
Recurrent 
Number Coil100  Corel10K   Outext 
7x7 - 5 0.947 0.486 0.758 
7x7 - 10 0.96 0.4975 0.758 
7x7 - 15 0.969 0.497 0.788 
7x7 - 20 0.974 0.503 0.789 
7x7 - 25 0.972 0.495 0.786 
Table 4: accuracy results for SVM classifier on all datasets. 
Window Size − 
Recurrent 
Number 
     
Coil100 Corel10K 
         
Outext 
7x7 - 5 
0.913 0.4 
0.614 
7x7 - 10 
0.925 0.402 0.628 
7x7 - 15 
0.928 0.396 
0.639 
7x7 - 20 
0.932 0.402 0.64 
7x7 - 25 
0.937 0.41 0.635 
Table 5: accuracy results for NB classifier on all datasets. 
 
Window Size − 
Recurrent 
Number 
 
      
Coil100 Corel10K Outext 
7x7 - 5 0.911 0.441 0.669 
7x7 - 10 0.911 0.436 0.692 
7x7 - 15 0.932 0.438 0.699 
7x7 - 20 0.929 0.438 0.702 
7x7 - 25 0.928 0.442 0.699 
Table 6: accuracy results for DAC classifier on all 
datasets. 
  Outext Corel10K 
 TCM 41.96 % ± 2.5% 16.71% ± 0.1% 
 MTH 33.85 % ± 2.4% 23.68% ± 0.1% 
 CTM 33.65 % ± 1.8% 21.22% ± 0.2% 
 NRFU 39.83 % ± 2.5% 5.38% ± 0.0% 
 CMTH 42.34 % ± 2.6% 31.76% ± 0.1% 
 
SCH 64.24%± 2.5% 32.72%± 0.1% 
DAC 
Proposed 
Technique 
69.22%± 0.1% 43.91%± 0.1% 
 TCM 42.95 % ± 3.0% 20.58 % ± 0.1% 
 MTH 45.31 % ± 2.6% 33.02 % ± 0.1% 
 CTM 27.43 % ± 1.3% 21.15 % ± 0.2% 
 NRFU 31.51 % ± 1.9% 11.69 % ± 0.1% 
 CMTH 51.01 % ± 1.6% 40.61 % ± 0.2% 
 
SCH 77.08% ± 0.1% 40.96% ± 0.1% 
SVM 
Proposed 
Technique 
77.58% ± 0.1% 49.57% ± 0.1% 
 TCM 37.24 % ± 1.7% 23.62 % ± 0.2% 
 MTH 25.54 % ± 1.1% 24.68 % ± 0.2% 
 CTM 34.97 % ± 1.5% 23.06 % ± 0.2% 
 NRFU 41.01 % ± 1.5% 22.71 % ± 0.2% 
 CMTH 38.71 % ± 1.5% 34.22 % ± 0.2% 
 
SCH 61.08% ± 0.1% 34.28% ± 0.2% 
NB 
Proposed 
Technique 
63.12% ± 0.1% 40.20% ± 0.1% 
 TCM 64.63 % ± 2.5 18.33 % ± 0.1% 
 MTH 47.70 % ± 2.3% 17.23 % ± 0.1% 
 CTM 67.38 % ± 2.7% 19.69 % ± 0.1% 
 NRFU 60.81 % ± 2.6% 15.96 % ± 0.1% 
 CMTH 64.45 % ± 2.8% 22.94 % ± 0.1% 
 
SCH 76.32 % ± 0.1% 22.96% ± 0.1% 
DT 
Proposed 
Technique 
76.38% ± 0.1% 23.3% ± 0.1% 
Table 7: Results of the proposed descriptor with SCH and 
texton based features using ML classifiers. 
 

102 Informatica 45 (2021) 97–105 E. M. Martey et al.  
 
the recurrent transformation of 5 provided good results 
hence adopted the same for extracting features for this 
experiment. Again, the number of retrieved images from 
each dataset was 6, 12, 18, 24 and 30.  
To compute the weight values of CCH and SCH 
features, the α value starts with zero (0) and increases to 
one (1) with a step of 0.1. The notation [x: y] on the results 
chart has x as α weight while y represents the β weight 
used for the retrieval task. 
Figures 3–6 present the Precision performances of the 
proposed technique on Outext, Corel-10K, and COIL100 
and Batik datasets. 
Figures 3, 4, 5 and 6 present the precision values of 
the five retrieved values (6, 12, 18, 24 and 30). Generally, 
precision values for the top five retrieved values describe 
a gentle distribution curve for all datasets.  When the two 
weight values (α, β) are the same, the retrieval result 
shows that the visual effect is the best.  That is when α is 
0.5 and β is also 0.5, the proposed technique generally 
performs best. 
Figure 7 presents the proposed technique results 
together with SCH and textons implemented by Martey et 
al. [23] on Corel-10K. Both experiments used retrieval 
values of 10 to 100 with a step of 10.  
From Figure 7, the result of our proposed scheme 
outperforms TCM, MTH, CTM, NRFUCTM, CMTH and 
 
Figure 3: Precision for various retrieved values for the Outext dataset. 
 
Figure 4: Precision for various retrieved values for Corel dataset. 
 
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
5 [ 0 . 0 : 
1 . 0 ]
5 [ 0 . 1 : 
0 . 9 ]
5 [ 0 . 2 : 
0 . 8 ]
5 [ 0 . 3 : 
0 . 7 ]
5 [ 0 . 4 : 
0 . 6 ]
5 [ 0 . 5 : 
0 . 5 ]
5 [ 0 . 6 : 
0 . 4 ]
5 [ 0 . 7 : 
0 . 3 ]
5 [ 0 . 8 : 
0 . 2 ]
5 [ 0 . 9 : 
0 . 1 ]
5 [ 1 . 0 : 
0 . 0 ]
Precision
Number of retrieved images
Accuracy R=
6
Accuracy R=
12
Accuracy R=
18
Accuracy R=
24
Accuracy R=
30
0
0,05
0,1
0,15
0,2
0,25
0,3
0,35
0,4
0,45
0,5
5 [ 0 . 0 : 
1 . 0 ]
5 [ 0 . 1 : 
0 . 9 ]
5 [ 0 . 2 : 
0 . 8 ]
5 [ 0 . 3 : 
0 . 7 ]
5 [ 0 . 4 : 
0 . 6 ]
5 [ 0 . 5 : 
0 . 5 ]
5 [ 0 . 6 : 
0 . 4 ]
5 [ 0 . 7 : 
0 . 3 ]
5 [ 0 . 8 : 
0 . 2 ]
5 [ 0 . 9 : 
0 . 1 ]
5 [ 1 . 0 : 
0 . 0 ]
Precision
Number of retrieved images
Accuracy R=
6
Accuracy R=
12
Accuracy R=
18
Accuracy R=
24
Accuracy R=
30

Effective Image Representation using Double Colour Histograms for ... Informatica 45 (2021) 97–105 103 
 
SCH used to index the Corel 10K. The experimental result 
shows improved retrieval performance of the proposed 
technique compared with the SCH and texton approaches 
for extracting image texture features. SCH and Textons 
were chosen because literature has demonstrated that they 
are very effective for extracting image textural 
information [27, 35, 39] and hence can be used as a 
benchmark for evaluating similar feature extraction 
schemes. The result demonstrates the proposed technique 
leverage the transformational (rotational, scaling, 
translation and deformation) invariant nature of the SCH 
and CCH. 
4 Conclusion 
This paper presents a double colour histogram that 
effectively improves image feature representation for 
colour and texture images in a given database. The present 
descriptor combines SCH and CCH into a single vector for 
indexing images. The indexing scheme used a vector 
dimension of 128 and was very effective for image 
retrieval. The descriptor demonstrates high robust feature 
set used to represent the colour and texture of the image. 
The final experiments were carried out over four public 
database sets. Three image databases, mainly from Batik, 
Coil100 and Outext, were used to assess the colour and 
texture representation capabilities of our proposed 
 
Figure 5: Precision for various retrieved values for Coil dataset. 
 
Figure 6: Precision for various retrieved values for Batik dataset. 
0
0,2
0,4
0,6
0,8
1
1,2
1 0 [ 0 . 0 : 
1 . 0 ]
1 0 [ 0 . 1 : 
0 . 9 ]
1 0 [ 0 . 2 : 
0 . 8 ]
1 0 [ 0 . 3 : 
0 . 7 ]
1 0 [ 0 . 4 : 
0 . 6 ]
1 0 [ 0 . 5 : 
0 . 5 ]
1 0 [ 0 . 6 : 
0 . 4 ]
1 0 [ 0 . 7 : 
0 . 3 ]
1 0 [ 0 . 8 : 
0 . 2 ]
1 0 [ 0 . 9 : 
0 . 1 ]
1 0 [ 1 . 0 : 
0 . 0 ]
Precision
Number of retrieved images
Accuracy R=
6
Accuracy R=
12
Accuracy R=
18
Accuracy R=
24
Accuracy R=
30
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
5[0.0 :
1.0]
5[0.1 :
0.9]
5[0.2 :
0.8]
5[0.3 :
0.7]
5[0.4 :
0.6]
5[0.5 :
0.5]
5[0.6 :
0.4]
5[0.7 :
0.3]
5[0.8 :
0.2]
5[0.9 :
0.1]
5[1.0 :
0.0]
Precision
Number of retrieved images = 6

104 Informatica 45 (2021) 97–105 E. M. Martey et al.  
 
descriptor. Corel10K was used to assess its ability to 
represent heterogeneous images. Experimental results 
have shown that our proposed method has strong texture 
and colour discrimination power and outperforms SCH 
and well-established texton-based features.  
Funding statement: This study is supported by the 
Provincial Science and Technology Key Research & 
Development Sector (2019). “Establishment and practice 
of cancer patient assessment assistant intelligent decision 
system” (Grant No. 19ZDYF0794). 
Conflicts of interest: The authors declare no conflict 
of interest concerning this work. 
References 
[1] T. Deselaers, D. Keysers and H. Ney. Features for 
image retrieval: an experimental comparison. Inf. 
Retrieval 11(2):77–107, 2008. 
[2] G. Pass and R. Zabith . Histogram refinement for 
content-based image retrieval, in Proceedings of 
IEEE Workshop on Applications of Computer 
Vision, 96–102, 1996.  
[3] P. Srivastava, N. T. Binh, and A. Khare. Content-
based image retrieval using moments. 
In International Conference on Context-Aware 
Systems and Applications, 228-237, 2013. 
[4] J. Huang, S. Kumar, M. Mitra, W. J. Zhu and R. 
Zabih. Image indexing using colour correlograms. In 
Proceedings of IEEE computer society conference 
on Computer Vision and Pattern Recognition, 762-
768, 1997. 
[5] V. Tyagi . Content-based image retrieval: ideas, 
influences, and current trends. Springer Nature, 133-
159, 2018. 
[6] A. Rao, R. K. Srihari and Z. Zhang. Spatial colour 
histograms for content-based image retrieval. In 
Proceedings 11th International Conference on Tools 
with Artificial Intelligence, 183-186, 1999. 
[7] C. Kurtz, A. Depeursinge, S. Napel, C. F. Beaulieu 
and D. L. Rubin.  On combining image-based and 
ontological semantic dissimilarities for medical 
image retrieval applications.  Medical image 
analysis, 18(7):1082-1100, 2014. 
[8] M. Tuceryan and A. K. Jain. Texture analysis. 
Handbook of pattern recognition and computer 
vision,  235-276, 1993. 
[9] D. K. Park, Y. S. Jeon and C. S. Won. Efficient use 
of local edge histogram descriptor. In Proceedings of 
the 2000 ACM workshops on Multimedia, 51-54, 
2000. 
[10] T. M. Ojala and T. M. Pietikainen. Multiresolution 
gray-scale and rotation invariant texture 
classification with local binary patterns”. IEEE 
Trans. Pattern Anal. Mach. Intell,  24: 971–987, 
2002. 
[11] S. Marula, R.P. Maheshwari and R. 
Balasubramanian. Local tetra pattern: A new feature 
descriptor for content-based image retrieval. IEEE 
transactions on image processing, 21(5): 2874–2886, 
2012. 
[12] R. M. Haralick, K. Shanmugam and I. Dinstein, 
Textural Features for Image Classification, In IEEE 
Transactions on Systems, Man, and Cybernetics, 
3(6):610-621, 1973. 
https://doi.org/10.1109/TSMC.1973.4309314 
[13] D. Zhang, A. Wong, M. Indrawan and G. Lu. 
Content-based image retrieval using Gabor texture 
features. IEEE Transactions Pami,  3656: 13âĂŞ15. 
2000. 
[14] B. S. Manjunath and W. Y. Ma. Texture features for 
browsing and retrieval of image data. IEEE 
 
Figure 7: Performance of proposed technique compared with SCH and textons on Corel-10K. 
0
0,05
0,1
0,15
0,2
0,25
0,3
0,35
0,4
10 20 30 40 50 60 70 80 90 100
PRECISION
NUMBER OF RETRIEVED IMAGES
NRFUCT
M
TCM
CTM
MTH
CMTH
SCH
Proposed
Fusion

Effective Image Representation using Double Colour Histograms for ... Informatica 45 (2021) 97–105 105 
 
Transactions on Pattern Analysis & Machine 
Intelligence,   18(8): 837–842, 1996. 
[15] I. J. Sumana, M. M. Islam, D. Zhang and G. Lu. 
Content-based image retrieval using curvelet 
transform. In 2008 IEEE 10th workshop on 
multimedia signal processing,   11-16, 2008. 
[16] F.S. Cohen, Z. G. Fan, and M. A. Patel. 
Classification of rotated and scaled textured images 
using Gaussian Markov random field models. IEEE 
Transactions on Pattern Analysis & Machine 
Intelligence,   13(02):192–202, 1991. 
[17] B. Julesz. Texton gradients: The texton theory 
revisited. Biological Cybernetics, 54(4):  245-251, 
1986. 
[18] G. H. Liu, L. Zhang, Y. K. Hou, Z. Y.  Li and J. Y. 
Yang. Image retrieval based on multi-texton 
histogram. Pattern Recognition,  43(7): 2380-2389, 
2010. 
[19] G. H. Liu and  J. Y. Yang. Image retrieval based on 
the texton co-occurrence matrix. Pattern 
Recognition,  41(12): 3521-3527, 2008. 
[20] Y. S. Kumari, V. V. Kumar and C. Satyanarayana. 
Texture classification using complete texton matrix. 
International Journal of Image, Graphics and Signal 
Processing,  9(10):  60, 2017. 
[21] B. Khaldi, O. Aiadi and K. M. Lamine. Image 
representation using complete multi-texton 
histogram. Multimedia Tools and Applications, 
79(11): 8267-8285, 2020. 
[22] V. Tyagi. Content-Based Image Retrieval Using a 
Short Run Length Descriptor. In Content-Based 
Image Retrieval; Springer: Singapore, 241–256, 
2017. 
[23] E. M Martey., H. Lei, X.  Li and O. Appiah. Image 
Representation Using Stacked Colour Histogram. 
Algorithms,  14(8): 228, 2021. 
[24] G. H. Liu, Z. Y. Li,  L. Zhang and Y. Xu. Image 
retrieval based on the micro-structure descriptor. 
Pattern Recognition.   44(9): 2123–2133, 2011. 
[25] G. H. Liu, J. Y. Yang and Z. Li. Content-based image 
retrieval using computational visual attention model. 
Pattern Recognition, 48(8):2554–2566, 2015. 
[26] G.H.  Liu and J.Y. Yang. Image retrieval based on 
the texton co-occurrence matrix.  Pattern Recogn.  
41(12):3521–3527, 2008. 
[27] G.H .Liu, and J.Y. Yang. Content-based image 
retrieval using colour difference histogram. Pattern 
Recognition, 46(1): 188–198, 2013. 
[28] W. Xingyuan and W. Zongyu. A novel method for 
image retrieval based on structure elements 
descriptor. Journal of Visual Communication and 
Image Representation, 24(1): 63–74, 2013. 
[29] X. Tian L. .Jiao, X. Liu and X. Zhang. Feature 
integration of eodh and colour-sift: Application to 
image retrieval based on codebook, Signal 
Processing: Image Communication,  29( 4): 530–
545, 2014. 
[30] S. R. Dubey,  S. K Singh and R. K Singh. Rotation 
and scale-invariant hybrid image descriptor and 
retrieval, Computers & Electrical Engineering, 46: 
288–302, 2015. 
[31] A. Nazir, and N.  Kashif. An efficient image retrieval 
based on fusion of low-level visual features, arXiv 
preprint, 2018.   
https://arxiv.org/abs/1811.12695  
[32] B.  Zafar, R. Ashraf., N. Ali, M. Ahmed, S. Jabbar 
and S. A. Chatzichristofis. Image classification by 
addition of spatial information based on histograms 
of orthogonal vectors, PloS one, 13(6): e0198175, 
2018. 
[33] E. M. Martey, Lei H., X Li., O. Appiah and N. S. 
Awarayi. Evaluation of RGB Quantization Schemes 
on Histogram-Based Content-Based Image 
Retrieval. In International Conference on Artificial 
Intelligence and Security, 736-747, 2021.  
[34] O.Appiah, E. M. Martey and E. Quayson. Effect of 
Window's Shape on Median Filtering.  In 
Proceedings of the 2019 IEEE AFRICAN, Accra, 
Ghana, 1–8, 2019. 
[35] A. E. Minarno, Y. Munarko, F. Bimantoro, A. 
Kurniawardhani and N. Suciati. Batik image 
retrieval based on enhanced micro-structure 
descriptor. In Proceedings of the 2014 Asia-Pacific 
Conference on Computer-Aided System 
Engineering (APCASE), South Kuta, Indonesia, 65–
70, 2014. 
[36] G. Mori, S. Belongie and J. Malik. Shape contexts 
enable efficient retrieval of similar shapes. In 
Proceedings of the 2001 IEEE Computer Society 
Conference on Computer Vision and Pattern 
Recognition. Kauai, HI, USA, 1, 2001. 
[37] B. Khaldi, O. Aiadi and M. L. Kherfi. Combining 
colour and grey-level co-occurrence matrix feature: 
A comparative study. IET Image Process, 13:1401–
1410, 2019. 
[38] Bala A. and Kaur, T. Local texton XOR patterns: A 
new feature descriptor for content-based image 
retrieval. Eng. Sci. Technol. Int. J., 19:101–112, 
2016.     
[39] M. E. Mensah, X. Li, H. Lei, A. Obed and N. C. 
Bombie. Improving Performance of Colour-
Histogram-Based CBIR Using Bin Matching for 
Similarity Measure. In International Conference on 
Artificial Intelligence and Security, 586-596, 2020. 
 
  

106 Informatica 45 (2021) 97–105 E. M. Martey et al.