www.ijecs.in International Journal Of Engineering And Computer Science ISSN:2319-7242 Volume 4 Issue 2 February 2015, Page No. 10310-10312 A Comprenssive Survey On Different Methods For Detecting Salient Object Manish M. Paliwal1, Dr. Prashant N. Chatur2 1 PG Scholar, Government College of Engineering, Department of Computer Science and Engg., Amravati. INDIA paliwalmanish1@gmail.com 2 HOD, Government College of Engineering, Department of Computer Science and Engg., Amravati. INDIA chatur.prashant@email.com Abstract: This survey paper consists of various image processing techniques used for the detection of the salient region. The paper also includes the comparative study of these techniques like precession recall analysis and the relative advantages and disadvantages. The major topic of our interest is the Salient Region of the object which is an important factor in most of the areas like cognitive psychology and neurobiology. The various technique are different in there logic and performance the involved. Comparing to all the techniques on there both positive and negative sides the paper presented here is the best in my behalf to its domain knowledge. Keywords: GMM , Spatial Distribution , Rarity , Visual Uniqueness etc. 1. Introduction In image processing domain a large number of problem domains are available. One of area of our interest is Detection of Salient Region in the given image. In the image processing operation the saliency can be considered as the root problem because it affects the large number of the natural problems. When we observe some scenes then we noticed that during taking the picture the human focused on the some particular object mostly compared to the other objects in similar image. The particular object is called as Salient object where the human put its most of the attention. Sometime it is also called the object of interest. The researchers devoted a lot of time on various techniques in this area. But still no method can achieve the full optimum performance what we required. Some methods performs well on their execution time but compromises with the available space. Some methods restricted in the shape and size of the image. A detailed survey described below in this paper. 2. Different Methods A. Soft Image Abstraction Technique The Soft Image abstraction technique is one of the best available techniques to detect the salient object inside the image. The main idea behind this method is to decompose the given image into various small regions which bear some spatial relationship with respect to each other. Once we have decomposed the image the next phase is to assign the global Saliency values to the image pixels. The assignment of the image saliency values is done on the basis of the weighted color contrast with respect to the all other components. For the sake of simplicity and better precision we took the concept of the Gaussian Mixer Model. GMM define the probability of the pixel belonging to the one cluster. The clustering achieved through the Orchard and Bauman’s Algorithms which initially starts with one cluster and iteratively uses the eigenvalues and eigenvector of the covariance matrix to specify the cluster splitting point. The Clustering divide the color channel into 12 different parts and choose only those color which comprises approx. 95% of the image pixels. In Soft image abstraction method the probability of the image color Ix belonging to the component c is given by the following formula(1) Where Wc,µc and ∑c indicate the weight, mean color and the covariance matrix of the cth component. The method divide the color channel into 12 different band so that number of color should be minimum from computational point of view. Also it takes only those image color which account a minimum of 95% of the whole image. The idea comes from the fact that not all color contribute to the image formation so take only relevant colors. Now the spatial distribution among the two GMM components can be found by the correlational formula given as follow(2) After computing the correlational matrix it is necessary to perform the clustering operation on the available component. Such clustering can be achieved by the message passing clustering because it does not require the predefined cluster numbers. When we perform the clustering then the probability of each pixel color Ix belonging to the cluster C becomes- Manish M. Paliwal1 IJECS Volume 4 Issue 2 February, 2015 Page No.10310-10312 (3) Page 10310 Where Ib is the quantized histogram bin color of Ix and C denotes the cluster. We know that high contrast is a good indication of the saliency so the visual uniqueness of the image component can be computed by the following formula- (4) Where D(ci,cj) is the spatial distance between centroid of the two GMM component ci and cj and the value of σ2 is taken as 0.4. The experimental result shows that the soft image abstraction method capable to reduce the mean absolute error by 25%. The method is also better in terms of the execution time. B. Salient Region Detection using Saliency Filters This method is proposed by the Federico Perazzi and et al. The concept of their method is somewhat similar to the previous methods but it was new as it makes use of the Gaussian Filter for the Contrast and Saliency estimation. The methods consist mainly four basics steps. In first step it decomposes the given image into compact, perceptually homogeneous element and abstract out the unnecessary details which are irrelevant for the further analysis. Based on this they check for the two important features in the image element which are pixel contrast and spatial distribution of the image pixel. The pixel contrast value determines the rate of uniqueness of each pixel. During the first phase of the method the main task is to keep the decomposed element as small as they could preserve the relevant structure and and the same time abstract out unnecessary details.. This type of decomposition is mainly achieved by the edge preserving localized over segmentation based on color. Second phase of the method is to compute the two measure of the contrast. For that they compare each of the image pixels from the other pixel to find the feature called “rarity”. As we commonly know that the image pixel of the background are highly distributed on the entire image compare to the foreground image pixels but it doesn’t meant that they can be consider the salient. So it is necessary that some large scale image segmentation techniques should be adopted to get this property. The rarity of the image pixel is computed by the following way. (5) Where Ui = uniqueness of the pixel segment i Pi= Position of the ith pixel segment Ci= color of the segment i in CIELab The complexity of the above operation is O (N2) where N is the number of image segment. But making use of the Gaussian weight wij it is possible to reduce the complexity up to O(N). For element Distribution measure of a segment i they computed spatial variance Di of its color ci. Low variance indicates the compact object which is more salient than the spatially widely distributed element. The Di can be computed as follow- (6) Where µi = weighted mean position of the color ci. Again the equation of the spatial distribution has the quadratic complexity but by introducing the Gaussian function we can get the linear time complexity. Saliency assignment is the next step towards the combining both feature uniqueness and spatial distribution. They normalized the value of both factor in the range of (0…1). These two values are independent so we can compute the Saliency value Si for each element (7) For there experimental purpose they fixed the value of k as 6 which is a scaling factor for the exponential function. The final saliency value for image pixel Si is defined by the weighted linear combination of the saliency Sj of its surrounding image elements. (8) The method is good in overall performance tested over the best processing architecture. C. Salient Object Detection using Regional Contest Based algorithms The method was suggested by the Ming Mang Cheng et al. This method was based on the Histogram based contrast method. The idea was similar to the other salient object detection method but in differ in terms of saliency value calculation. The saliency of a pixel in the image is defined by the color contrast value of the all other pixel in the same image. The saliency value function for the calculation can be given as follow(9) Where D(Ik,Ii) is the color distance matrix between the pixels Ik and Ii in the image. Pixel with the same color has same saliency value by this equation. As the number of the image color are very large so the complexity of calculation is near O (N2) but it can be brought down to linear by reducing the number of image color. Instead of taking 256 different colors we quantized the color in the 12 different color bands. So it reduces the total available color. Again to get the better result we took only those colors which comprise the 95% of the overall image. The methods relay on the fact that the high contrast to ones surrounding region is stronger for the saliency of the region. For that purpose they introduce the region contrast (RC) method. They first segment the image into regions and then make a color histogram for the each region. The saliency value of each region is calculated as follow(10) Where w(ri) is the weight of region ri and Dr(.,.) is the color distance matrix between two region. For the space efficiency consideration we take the sparse histogram representation as the number of color in image region is very small. They also observe that the image region with a large boundaries overlapped by the other region are often treated as non-salient object. The performance of this method is compared against the available 15 different methods and found suitable for the detection. The HC method possesses the linear complexity. The variation of the HC is the RC method which is somewhat complex but yields a superior quality saliency maps. Manish M. Paliwal1 IJECS Volume 4 Issue 2 February, 2015 Page No.10310-10312 Page 10311 C. Supervised approach for salient object detection using condition random field This method was proposed by the Tie Liu et al. The idea of their method comprises of two steps. First they model the salient object detection problem by a condition random field (CRF) where they combine the multiple features through the CRF learning. They also propose a set of novel local, global salient feature to specify generic salient object. They initially provide the solution for the single image then they extended their idea for the sequential images. They define the problem as binary labeling task. The probability of a labeling configuration A={ax} is given by following conditional random function(11) Where Z is the partition function They defined the energy E(A|I) as follow- (12) Where λk is the weight of kth feature and x, x’ are two adjacent pixels Fk(ax,I) indicate probability of a pixel x belongs to the salient object. M. M. Paliwal received his B.E. degree in Computer Science and Engineering from University institute of Technology,RGPV, Bhopal, Madhya Pradesh, India in 2012, pursuing M.Tech degree in Computer Science and Engineering from Government College of Engineering, Amravati, Maharashtra, India. His research interest includes image processing, Algorithm Analysis, Operating System concept and Data mining. At present he is engaged in Salient Region detection Technique which will be giving a better performance than the existing Region Detection Techniques. Dr. P. N. Chatur has received his M.E degree in Electronics Engineering from Government College of Engineering Amravati, Maharashtra, India and Ph.D. degree from Amravari University. He has published twenty papers in international journals. His area of research includes Artificial Neural Network, Data Mining, Data Stream Mining and Cloud Computing. Currently, he is Head of Computer Science and Engineering & Electronics Engineering Department at Government College of Engineering Amravati, Maharashtra, India. At present he is engaged with large database mining analysis and stream mining. 3. Conclusion In this survey paper the aim is to study the different techniques available for the salient object detection on their both positive and negative sides. Some of the techniques are best in their complexity related issue but didn’t applicable on the all image size as they are restricted in the image window size. Some of the methods possess faster execution time on the all available system architecture. Beside this our aim is to direct the research of salient object detection in specific direction so that the difficulties which are still present can be solved. We hope that our review will direct the researcher to work well in this direction to achieve better efficient solution. References [1] Ming Ming Cheng and Jonathan Warrell , “Efficient Salient Region Detection with Soft Image Abstraction” IEEE International Conference on Comuter Vision,2013. [2] F. Perazzi, P. Kr¨ahenb¨uhl, Y. Pritch, and A. Hornung, “Saliency filters: Contrast based filtering for salient region detection”, In IEEE CVPR, pages 733–740, 2012. [3] Tie Liu, Zejian Yuan, “Learning to Detect a Salient Object,” IEEE Transactions on pattern Analysis and Machine Intelligence. vol. 33, February 2011. [4] Ming Mang Cheng, Guo-Xin Zhang, “Global Contrast based Salient Region Detection,” In IEEE CVPR,pages 409-416,2011. [5] Niloy J. Mitra, Xiaolei Huang, “Salient Object Detection and Segmentation,” IEEE Trans. On Pattern, Technical Report, TPAMI-2011-10-0753, 2011. Author Profile Manish M. Paliwal1 IJECS Volume 4 Issue 2 February, 2015 Page No.10310-10312 Page 10312