A COMPARATIVE ANALYSIS OF IMAGE SEGMENTATION TECHNIQUES

International Journal of Computer Engineering & Technology (IJCET) Volume 9, Issue 5, September-October 2018, pp. 64 69, Article ID: IJCET_09_05_009 Available online at http://www.iaeme.com/ijcet/issues.asp?jtype=ijcet&vtype=9&itype=5 Journal Impact Factor (2016): 9.3590(Calculated by GISI) www.jifactor.com ISSN Print: 0976-6367 and ISSN Online: 0976 6375 IAEME Publication A COMPARATIVE ANALYSIS OF IMAGE SEGMENTATION TECHNIQUES Sreedhar T Ph.D Research Scholar, Department of Computer Science, Erode Arts & Science College (Autonomous), Erode, Tamil Nadu State. Dr. Sathappan S Research Supervisor and Associate Professor, Department of Computer Science, Erode Arts & Science College (Autonomous), Erode, Tamil Nadu State. ABSTRACT Nowadays image processing occupies a dynamic role in every area. Image processing refers to progress the images and mine attributes from images for various process. Image segmentation is one of the processes of image processing which is used to segregate an image into different homogenous segments. It is an essential process of pixel clustering, partitions raw image into non-overlapping regions. Image segmentation plays an important role in almost all image processing applications. This paper focuses to provide an overview of the literature in image segmentation with various techniques implemented in them, their merits and demerits etc. Comparison based on accuracy was also done to prove the efficiency of the various techniques for image segmentation. The comparison results show the best image segmentation method among them. Key words: Image Partition, Image processing, Image Segmentation, Semantic Image Segmentation. Cite this Article: Sreedhar T and Dr. Sathappan S, A Comparative Analysis of Image Segmentation Techniques. International Journal of Computer Engineering and Technology, 9(5), 2018, pp. 64-69. http://www.iaeme.com/ijcet/issues.asp?jtype=ijcet&vtype=9&itype=5 1. INTRODUCTION The retrieval of valuable information from an image is one of the major goals of image processing [1]. The retrieval information is done in such a way that it will not affect the other features of that image. Image segmentation [2] is one of the essential steps of image processing. Image segmentation process splits an input image into multiple segments where each segment will represent a kind of information in the form of intensity, texture or color. Hence it is more important to isolate the boundaries of any image in the form of its segments. The process of image segmentation allocates a single value to each pixel of an image in order to make it easy to differentiate between different regions of any image. On the basis of http://www.iaeme.com/ijcet/index.asp 64 editor@iaeme.com

A Comparative Analysis of Image Segmentation Techniques three image properties such as texture, color and intensity of an image, the differentiation between image segments is done. The importance of image segmentation can t be neglected because it is utilized in various fields such as removing noise from an image, satellite imaging, medical images, computer vision, image retrieval, machine vision, military, biometrics, recognizing objects and extracting features from the given image. In this paper various image segmentation techniques are analyzed and compared in terms of accuracy. 2. LITERATURE SURVEY An Ontology Based Semantic Image Segmentation (OBSIS) [3] was projected for Image Segmentation and Object Segmentation. At first, a Dirichlet Procedure Mixture model was connected which changed over the low-level visual space into a moderate semantic space. This method reduced the spatial property of input options. At that point, numerous Conditional Random Fields (CRFs) was connected on the options which freely learned and separately weighted inside the situation. Consequently, the segmentation of pictures into objects was reduced to a classification task, where protest deduction was given a contribution to an Ontology Model. A holistic approach [4] was anticipated for scene understanding which reasons together about division, nearness, objects of a category in an image. The holistic issue was confined as a structure expectation issue in an exceeding graphical model delineated over hierarchies of regions of various sizes. Additionally, to the present auxiliary variables encoding the scene kind, the presence of a given category in a scene and the accuracy of the bounding boxes output by an object detector were outlined. Also, a shape prior was consolidated for objects with the very much characterized shape which appears as a delicate soft mask learned from coaching examples. A semantic segmentation approach [5] was planned for generating category specific image segmentation. Two novel highlights were presented that utilized the quantized information of the Discrete Cosine Transform (DCT) in a Semantic Texton Forest based framework (STF) for the method of semantic segmentation of images. It consolidated both the data of color and texture in images along with semantic segmentation. The proposed framework was intended to abuse the integral options in a economical manner. The DCT approach primarily based options outlined advanced textures that are represented in the frequency domain. This can be differing from the STF approach, it had been obtained straightforward textures by calculative the distinction between the intensity of pixels. Random forest algorithm is employed to feature the novel options within the STF system. The planned linguistics segmentation approach needed a restricted quantity of resources. A hybrid hierarchical model [6] was proposed for semantic picture segmentation. This model makes full utilization of casual and non-casual signals in common pictures by hybridizing the Hierarchical Conditional Random Field (HCRF) and Bayesian Network (BN) model. The HCRF was used to deliver beginning semantic sub-scene expectations by capturing the non-causal relationship such as inter class co-occurrence statistics and appearance features. The BN was utilized to display relevant associations for each semantic sub-scene as class measurements from its neighboring locales of which its restrictive probabilities were found out naturally from preparing information. In order to encipher the structure of the discourse dependencies for sub-scenes within the initial predictions, the learned BN structure was used. It created last refined expectations. Effective semantic picture division display [7] was proposed with multi-class positioning earlier. With a specific end goal to fulfill the equivocalness of appearance comparability, the planned model used most extreme edge based Structural Support Vector Machine (S-SVM) by consolidating different level of signals. At that point, a novel multi-class positioning based http://www.iaeme.com/ijcet/index.asp 65 editor@iaeme.com

Sreedhar T and Dr. Sathappan S global constraint was proposed to hold the object classes and planned to carry the thing categories to be assumed once labeling regions within an image. This methodology was additional balanced between communicatory power for heterogeneous regions and potency of looking exponential space of conceivable label combinations. An inter-class co-occurrence statistic was introduced as a pairwise constraint to alter a joint deduction of naming inside a picture with better consistency. Then mix with the prediction from global and local cues signs in view of S-SVMs structure. A robust modified Gaussian mixture model with rough set [8] was anticipated for image segmentation. With a specific end goal to make the Gaussian mixture model a lot of robust to noise, a brand-new spatial weight issue was created. The created spatial weight issue utilized the computation of pixels in its immediate neighborhood. At that point the over smoothness was reduced by introducing a unique previous issue. It joined the spatial data among neighborhood pixels. Each Gaussian segment was characterized through the automated determination of rough regions and in keeping with this the posterior likelihood of every component was calculated with reference to the region it locates. An end-to-end trainable Deep Convolutional Neural Network (DCNN) [9] was introduced for semantic segmentation of pictures. Here, a technique was planned to counter the result of loss of effective spatial resolution washed out high-frequency details and ends up in hazy protest limits. In this technique, the semantic segmentation was joined with the semantically sophisticated edge detection that makes category boundaries specific within the model. A relatively easy, memory economical model was made by adding boundary detection to the SEGNET encoder-decoder design. Notwithstanding that, boundary detection was conjointly enclosed in FCN sort models and got wind of a high finish classifier ensemble. A novel Deep Parsing Network (DPN) [10] was planned for semantic image segmentation. At first, DPN bound together the inference and learning of unary term and pairwise terms in a very single convolutional network. Throughout the back propagation of DPN, no reiterative inference was needed. Then high order relations and mixtures of label contexts were incorporated to its pairwise terms modeling. At long last, DPN was based upon convolutional activities on Convolutional Neural Network (CNN) therefore simple to be parallelized and accelerate the picture segmentation process. A deep architecture [11] was planned for real time segmentation of pictures. The core of this architecture was ConvNet that could be a novel layer that used residual connections and factorized convolutions with a specific end goal to remain exceptionally proficient while as yet holding momentous execution. The residual connections supposed a breakthrough as a result, they avoid the degradation issue that is available in models with a lot of stacked layers. This has allowed terribly deep networks (with many layers) to accomplish exceptional characterization execution. A more extensive engineering was utilized which influences a to a great degree proficient utilization of its limited reduced quantity of layers to attain correct segmentation in real time. A contextual framework called as Contextual Hierarchical Model (CHM) [12] was proposed for semantic image segmentation. For semantic segmentation, the planned model learned discourse info in an exceedingly hierarchic framework. A classifier was trained at every level of hierarchy that was supported down sampled input pictures and outputs of previous levels. Then, the ensuing multi resolution contextual info was combined with a classifier to section the input image at original resolution. This coaching strategy was allowed for joint posterior chance optimization at multiple resolutions through the hierarchy. http://www.iaeme.com/ijcet/index.asp 66 editor@iaeme.com

A Comparative Analysis of Image Segmentation Techniques 2.1. Comparison of Segmentation Techniques The image segmentation methods described in the above section is analyzed and compared based on methods used, their merits, demerits and the parameters used for experimental results. The comparison is shown in Table I. Table 1 Comparative Study of an Existing Methods S. No. Method Merits Demerits Results Failed to segment of touching Ontology Based Provide human like objects with similar features, 1 Semantic Image image labeling Not effective enough due to the Segmentation limited number of concepts 2 Holistic Approach Much Faster 3 4 5 6 7 Semantic segmentation approach, Discrete Cosine Transform, Semantic Texton Forest based framework Hierarchical Conditional Random Field, Bayesian Network Structural Support Vector Machine Robust modified Gaussian mixture model with rough set Deep Convolutional Neural Network 8 Deep Parsing Network 9 Deep architecture 10 Contextual Hierarchical Model Required a limited amount of resources Provide more accurate and reliable labeling results Provide highly competitive performance Improve the segmentation accuracy Achieves better performance Speed up image segmentation process Used minimized number of layers to achieve accurate segmentation in real time General labeling model with high accuracy Varying threshold value affect the segmentation accuracy Large number of trees may make the algorithm slow for real time predictions Considering only a limited number of features for segmentation Failed to integrate multi-source cues which contain complementary information for image segmentation Use of spatial information in mixture models becomes more complex Excessive size and complexity Some objects in images cannot be localized accurately Even though better mean Intersection-over- Union (IoU) and category IoU might seem like low values It can be prone to error due to absence of any global constrains Global Accuracy = 86% Global Accuracy = 80.6% Accuracy = 76.35% Global accuracy = 82.2% Global accuracy = 77.8% PRI = 0.9885 Accuracy (JS values) = 0.8908 Overall Accuracy = 90.3% Mean boundary accuracy = 93.3% Accuracy = 92.8% F- value = 86% G-mean = 92.48% Pixel Accuracy = 95.37% In Table I, the different image segmentation approaches are analyzed based on accuracy. The Contextual Hierarchical Model [12] has better accuracy of 95.37% than the other methods. http://www.iaeme.com/ijcet/index.asp 67 editor@iaeme.com

Accuracy Sreedhar T and Dr. Sathappan S 3. PERFORMANCE EVALUATION OF EFFICIENT METHODOLOGIES The performance of the economical methodologies within the literature are analyzed and compared among them to see the comparative performance potency. The methods considered for the analysis are Ontology Based Semantic Image Segmentation [3], Holistic Approach [4], Semantic segmentation approach, Discrete Cosine Transform, Semantic Texton Forest based framework [5], Hierarchical Conditional Random Field, Bayesian Network [6], Structural Support Vector Machine [7], Robust modified Gaussian mixture model with rough set [8], Deep Convolutional Neural Network [9], Deep Parsing Network [10], Deep architecture [11] and Contextual Hierarchical Model [12]. The comparison is carried out by evaluating the methods in terms of accuracy. 3.1. Accuracy Accuracy is defined as the understanding of a measurement to the true value. It is given as (1) The problem is to find the optimal configuration of parameter setting given in the Literature Survey. Figure 1 compares the various method, CHM has the better accuracy than the other methods. Accuracy and Methods are plotted in X axis and Y axis respectively. Thus, the comparative analysis in the Figure 1 prove that CHM has higher accuracy rate than the other. In general CHM has the labeling models with the high accuracy. Accuracy 100 80 60 40 20 0 Methods Ontology Based Semantic Image Segmentation Holistic Approach Semantic segmentation approach, Discrete Cosine Transform, Semantic Texton Forest based framework Hierarchical Conditional Random Field, Bayesian Network Structural Support Vector Machine Robust modified Gaussian mixture model with rough set Deep Convolutional Neural Network Deep Parsing Network Deep architecture Contextual Hierarchical Model Figure 1 Comparative Study of an Existing Methods http://www.iaeme.com/ijcet/index.asp 68 editor@iaeme.com

A Comparative Analysis of Image Segmentation Techniques 4. CONCLUSIONS Image segmentation is considered to be the recent ever-growing processes that focus on attaining higher values of accuracy in order to support various applications. Here this paper provides the recent developments in the image segmentation techniques are analyzed by describing the novel ideas incorporated in them. The analysis of these schemes provides better understanding of the steps involved in each process thus increasing the scope for finding the efficient techniques to achieve maximum accurate performance. The comparison of the efficient techniques is carried out in terms of accuracy. The survey concludes that the Contextual Hierarchical Model [12] was efficient among all the techniques in terms of accuracy. This survey also helps in deriving the motivation for our future researches as well. REFERENCES [1] Lei, T., & Udupa, J. K. (2003). Performance evaluation of finite normal mixture modelbased image segmentation techniques. IEEE Transactions on Image Processing, 12(10), 1153-1169. [2] Khan, W. (2013). Image Segmentation Techniques: A Survey. In Journal of Image and Graphics, 1(4), 166-170. [3] Zand, M., Doraisamy, S., Halin, A. A., & Mustaffa, M. R. (2016). Ontology-based semantic image segmentation using mixture models and multiple CRFs. IEEE Transactions on Image Processing, 25(7), 3233-3248. [4] Yao, J., Fidler, S., & Urtasun, R. (2012, June). Describing the scene as a whole: Joint object detection, scene classification and semantic segmentation. In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on (pp. 702-709). IEEE. [5] Ravì, D., Bober, M., Farinella, G. M., Guarnera, M., & Battiato, S. (2016). Semantic segmentation of images exploiting DCT based features and random forest. Pattern Recognition, 52, 260-273. [6] Wang, L. L., & Yung, N. H. (2015). Hybrid graphical model for semantic image segmentation. Journal of Visual Communication and Image Representation, 28, 83-96. [7] Pei, D., Li, Z., Ji, R., & Sun, F. (2014). Efficient semantic image segmentation with multiclass ranking prior. Computer Vision and Image Understanding, 120, 81-90. [8] Ji, Z., Huang, Y., Xia, Y., & Zheng, Y. (2017). A robust modified Gaussian mixture model with rough set for image segmentation. Neurocomputing, 266, 550-565. [9] Marmanis, D., Schindler, K., Wegner, J. D., Galliani, S., Datcu, M., & Stilla, U. (2018). Classification with an edge: improving semantic image segmentation with boundary detection. ISPRS Journal of Photogrammetry and Remote Sensing, 135, 158-172. [10] Liu, Z., Li, X., Luo, P., Loy, C. C., & Tang, X. (2015, December). Semantic image segmentation via deep parsing network. In Computer Vision (ICCV), 2015 IEEE International Conference on (pp. 1377-1385). IEEE. [11] Romera, E., Alvarez, J. M., Bergasa, L. M., & Arroyo, R. (2017, June). Efficient convnet for real-time semantic segmentation. In Intelligent Vehicles Symposium (IV), 2017 IEEE (pp. 1789-1794). IEEE. [12] Seyedhosseini, M., & Tasdizen, T. (2016). Semantic image segmentation with contextual hierarchical models. IEEE transactions on pattern analysis and machine intelligence, 38(5), 951-964. http://www.iaeme.com/ijcet/index.asp 69 editor@iaeme.com