Digital Image Processing
Digital Image Processing
ABSTRACT
Emotion in natural scene images plays an important role in the way humans perceive an image. Based on
the emotion (happiness, sadness, fear, anger etc.) of any human being the images that are viewed by that person can
have a significant impact in a sense that if the person is for example in happy mood and he/she views an image that
is pleasing then he/she would have a better sense of attachment towards that image and would not accept an image
that depicts sadness as an emotion. Although different people may interpret the same image in different ways, we
still can build a universal classification for different emotions. Thise Emotion Detection from Natural Scene Images
is a new concept in an innovative field of Image Processsing domain. Image processing domain has always proven
to be a challenging criterion in field of research and development.
Keyword: - Image Description, Image classification, Content Based Image Retrieval (CBIR).
1. INTRODUCTION
The problem of emotion detection poses interesting questions from a research point of view; for instance:
how to model the text for the detection task, what features offer the best prediction/detection power, and to what
extent it is even possible to accurately distinguish subjective labels such as emotions from a given source text [2]
[3]. To predict emotion, we carry out a fairly traditional machine learning method with the addition of feature
selection techniques. Specifically, the experiments here use a set of six basic emotions [2].
2. CLASSIFICATION
We used two methods for the classification of the local semantic concepts, k-Nearest Neighbour and
Support Vector Machine classifiers. Same classification methods were used in the initial method.
6. IMAGE SEGMENTATION
Automatic image segmentation is a difficult task. A variety of techniques have been proposed in the past,
such as curve evolution, energy diffusion, and graph partitioning. Many existing segmentation techniques work well
for mages that contain only homogeneous colour regions, such as direct clustering methods in colour space. These
apply to retrieval systems working only with colours [1].
9. ALGORITHMS
9.1 Color Moment
Color moments are measures that can be used differentiate images based on their features of color. Once
calculated, these moments provide a measurement for color similarity between images. These values of similarity
can then be compared to the values of images indexed in a database for tasks like image retrieval. The basis of color
moments lays in the assumption that the distribution of color in an image can be interpreted as a probability
distribution. Probability distributions are characterized by a number of unique moments (e.g. Normal distributions
are differentiated by their mean and variance). It therefore follows that if the color in an image follows a certain
probability distribution, the moments of that distribution can then be used as features to identify that image based on
color. Three central moments of a image's color distribution. They are Mean, Standard deviation and Skewness. A
color can be defined by 3 or more values. (Here we will restrict ourselves to the HSV scheme of Hue, Saturation and
brightness, although alternative encoding could just as easily be used.) Moments are calculated for each of these
channels in an image. An image therefore is characterized by 9 moment’s 3 moments for each 3 color channels.
……………………(1)
The 'value' of the image originally referred to the gray scale value of the specified pixel. The value could be
anything, from a binary on/off value to 32-bit color and beyond [3]. Note that 32-bit colour will yield a 232 *232 co-
occurrence matrix. Really any matrix or pair of matrices can be used to generate a cooccurrence matrix, though their
main applicability has been in the measuring of texture in images, so the typical definition, as above, assumes that
the matrix is in fact an image. It is also possible to define the matrix across two different images. Such a matrix can
then be used for colour mapping. Note that the (Δx, Δy) parameterization makes the co-occurrence matrix sensitive
to rotation. We choose one offset vector, so a rotation of the image not equal to 180 degrees will result in a different
co-occurrence distribution for the same (rotated) image. This is rarely desirable in the applications co-occurrence
matrices are used in, so the co-occurrence matrix is often formed using a set of offsets sweeping through 180
degrees (i.e. 0, 45, 90, and 135 degrees) at the same distance to achieve a degree of rotational invariance [5].
Table-1: Confusion Matrix of the SVM Concept Classification (C=8, g=0.125). Classification is in %
Overall w g r s s f f t
67,8%
water 71,8 1,4 7,0 1,4 14,1 4,2 0,0 0,0
grass 9,2 40,0 7,5 0,8 0,0 18,3 23,3 0,8
rock 7,4 0,5 77,5 6,9 0,5 2,0 2,5 2,9
sand 0,0 6,7 23,3 53,3 10,0 3,3 0,0 3,3
sky 8,0 0,0 0,9 1,8 82,1 2,7 4,5 0,0
foliage 3,1 14,8 4,8 0,0 0,0 71,6 3,9 1,7
flowers 0,0 11,6 2,0 1,5 0,0 17,2 67,7 0,0
trunks 1,6 3,1 23,4 7,8 0,0 9,4 1,6 53,1
Precision 54,26 43,24 75,24 38,109 86,7 69,2 73,631 73,9
descriptor with 64 bins results. In the higher scale, as no image division is done because only the most important
image features will be present, 4 bins result. A total of 68 bins are used to describe the image in scale-space. Images
will be compared using the Euclidean distance between histograms. The provided results will be compared with the
ones that result from the use of the histogram in the low scale only. Improved classification using the nearest class
mean and neural networks will be used. A higher level semantic annotation, based on this low level descriptor that
results from the multiscale image analysis, will be extracted [4].
10. RESULTS
This section summarizes the results of the proposed approach as tabulated in folloeing Table-2.
Color 52,3%
Co-occurrence matrix 41,2%
Gabor feature 43,4%
Edge direction 25,3%
Color+Co-ocurance matrix 59,8%
Color+Gabor feature 62,5%
Color+Edge direction 56,7%
All features 67,8%
We measured the quality of human body detection by comparing the obtained results with manual
detection. We calculated the overlap and left-out feature. Overlap feature determines what percentage of the manual
detection (MD) is covered by the obtained result (OR).
Overlap=area (OR\MD) area (MD)
Left-out feature determines what percentage of the obtained result is not covered by the manual detection.
Le ft out = area (OR MD) area (OR)
Our method for human body detection was tested on 15 images and we achieved average Overlap 92;
06% and average Left-out 15; 42%. The method works well if the person is standing straight. It is a typical pose on
holiday pictures. If person is sitting or lying some errors may occur. (See Figure 6) As a next step we tested which
low level features are most relevant in classification Process. Results obtained using SVM classifier can be find in
Table 2. It is obvious that color feature give a good result, but its combination with texture feature leads in even
better accuracy.
The ground truth for sub region membership to one of the eight semantic concepts was annotated manually. To-
gather we annotated 1028 sub regions. The class sizes vary from 54 (trunks) up to 192 (sky), because sky appears
more often in the images than trunks. The classifiers are challenged with the inequality in the class sizes and the
visual similarity of image regions that belong to different classes. The Table 3 shows that the SVM classification
performs better than the KNN classification. We can see a correlation between the class size and the classification
result. Sky, foliage, and rocks are the largest classes and they are also classified with the highest accuracy. In Table
1 is displayed confusion matrix of the SVM concept classification.
Class Classification
size Accuracy Accuracy
KNN SVM
Sky 192 77,2 % 82,1 %
Water 139 53,4 % 71,8 %
Grass 111 20,7 % 40,0 %
Trunks 54 43,8 % 53,1 %
Foliage 166 66,7 % 71,6 %
Sand 103 47,6 % 53,3 %
Rocks 171 66,0 % 77,5 %
Flowers 94 57,7 % 67,7 %
11. CONCLUSION
Emotion Detection from Natural Scene Images is a new concept in an innovative field of Image
Processing domain. Image processing domain has always proven to be a challenging criterion in field of research
and development.
REFERENCES
[1] A. Mojsilovic and B. Gomes, J.and Rogowitz, “Semantic-Friendly Indexing and Quering of Images Based on the
Extraction of the Objective Semantic Cues” International Journal of Computer Vision, 56:79–107, 2004.
[2] E. Sikudova´, “On Some Possibilities of Automatic Image Data Classification” PhD thesis, Comenius
University, Bratislava, Slovakia, March 2006.
[3] Jianjiang Lu, Zhenghui Xie, Ran Li, Yafei Zhang, and Jiabao Wang., “A Framework of Cbir System Based On
Relevance Feedback.
[4] J. Vogel and B. Schiele, “Semantic Modeling of Natural Scenes For Content-Based Image Retrieval. Int. J.
Comput. Vision” 72:133–157, April 2007.
[5] Y. Liu, D. Zhang, G. Lu, and W. Ma., “A Survey of Content-Based Image Retrieval With High-Level Seman-
Tics” Pattern Recognition, 40(1):262 – 282, 2007.
BIOGRAPHIES
1
Mr. Kuldip Kandekar pursuing Bachelor Degree in Computer Engineering from Sir Visvesvaraya Institute of Technology,
Chincholi, Nashik, Savitribai Phule Pune University, Maharashtra, India.
2
Mr. Ankush Lahamge pursuing Bachelor Degree in Computer Engineering from Sir Visvesvaraya Institute of Technology,
Chincholi, Nashik, Savitribai Phule Pune University, Maharashtra, India.
3
Mr. Ajit Tadvi pursuing Bachelor Degree in Computer Engineering from Sir Visvesvaraya Institute of Technology,
Chincholi, Nashik, Savitribai Phule Pune University, Maharashtra, India.
3
Miss. Abhilasha Karpe pursuing Bachelor Degree in Computer Engineering from Sir Visvesvaraya Institute of
Technology, Chincholi, Nashik, Savitribai Phule Pune University, Maharashtra, India.
4
Prof. P.S. Daware Assistant Professor in the Department of Computer Engineering, Sir Visvesvaraya Institute of
Technology, Chincholi, Nashik, Savitribai Phule Pune University, Pune, Maharashtra, India.