Approach For Image Data Mining Cultural Studies Essay


This paper presents a probablsitic approach for recognition of natural images. The paper implements the isolation of natural image based on color localization and shape with bounding regions. The extracted regions are classified using multiple features and train them for the recognition of image objects. With multiple features calculated for the recognition of natural image samples. This approach is developed with the focus of increasing the performance of the current recognition system. This proposed approach of mining based on the local feature content of natural image is evaluated for performance evaluation.

Keywords: local features, natural images, image retrieval system, probabilistic approach, retrieval accuracy.

I. Introduction

Past few years have seen many advanced techniques evolving in data mining systems. Applications like art, medicine, entertainment, education, manufacturing, etc. make use of vast amount of visual data in the form of images. This envisages the need for fast and effective retrieval mechanisms in an efficient manner.

A major approach directed towards achieving this goal is to use low-level visual features of the image data to segment, index and retrieve relevant images from the image database. Recent dataminig systems based on features like color, shape, texture, spatial layout, object motion, etc., are cited in [1], [2]. Of all the visual features, color is the most dominant and distinguishing one in almost all applications. Hence, our approach is to segment out prominent regions in the image based on color and pick out their features. We then use shape features of these regions to obtain shape index used for retrieving based on shape matching.

Shape is an important feature for perceptual object recognition and classification of images. Shape description or representation is an important issue both in object recognition and classification. Many techniques such as chain code, polygonal approximations, curvature, fourier descriptors, radii method and moment descriptors have been proposed and used in various applications [3]. Recently, techniques using shape measure as an important feature have been used for CBIR. Features such as moment invariants and area of region have been used in [4], [5], but do not give perceptual shape similarity. Cortelazzo [6] used chain codes for trademark

image shape description and string matching technique. The chain codes are not normalized and string matching is not invariant to shape scale. Jain and Vailaya [7] proposed a shape representation based on the use of a histogram of edge directions. But these are not normalized to scale and computationally expensive in similarity measures. Mehrotra and Gary [8] used coordinates of significant points on the boundary as shape representation. It is not a compact representation and the similarity measure is computationally expensive. Jagadish [9] proposed shape decomposition into a number of rectangles and two pairs of coordinates for each rectangle are used to represent the shape. It is not rotation invariant. A region-based shape representation and indexing scheme that is translation, rotation and scale invariant is proposed by Lu and Sajjanhar []. It conforms to human similarity perception. They have compared it to Fourier descriptor model and found their method to be better. But, the images database consists of only 2D planar shapes and they have considered only binary images. Moreover, shapes with similar eccentricity but different shapes are retrieved as matched images. Our aim is to extend this method on color images and also to improve efficiency and effectiveness in retrieval. We segment out color image regions from images using dominant colors [11] and apply shape indexing to retrieve images based on shape features. Our shape indexing feature and similarity measure is different and shown to be effective in retrieval compared to the measure used in [10].

II.Feature representation

The initial step in our approach is to segment images into regions based on dominant colors []. Image regions thus obtained after segmentation are used as input to the shape module. The region-based shape representation proposed in [10] is modified to calculate the shape features required for our proposed shape indexing technique and similarity measure. It is simple to calculate and robust. We show that the retrieval effectiveness is better compared to the method in [10].

Color segmentation approach

To segment images based on dominant colors, a color quantization in RGB space using 25 perceptual color categories is employed [11]. From the segmented image we find the enclosing minimum bounding rectangle (MBR) of the region, its location, image path, number of regions in the image, etc., and all these are stored in a metafile for further use in the construction of an image index tree.

Color Space Categorization

The entire RGB color space is described using a small set of color categories that are perceptual to humans. This is summarized into a color look-up table as depicted in table 1. A smaller set is more useful since it gives a coarser description of the color of a region thus allowing it to remain same for some variations in imaging conditions. We have taken a table of 25 perceptual colors chosen from the standard RGB color palette table.

Table 1: Color look-up table

Color matching and region selection

The method relies on the fact that boundaries where perceptual color changes occur must be found before any cluster in color space can be interpreted as corresponding to a region in image space. The RGB color space is partitioned into subspaces called color categories. The perceptual color of a pixel can be specified by the color category into which it maps.

The procedure below segments the image into regions according to their perceived color. It involves mapping all pixels to their categories in color space, and grouping pixels belonging to same category. A color will be selected from 25 predefined colors which is very near to image pixel color and it will be stored as new color pixel in the image. Using p the image pixel value and C the corresponding entry in the color table, Color distance Cd is calculated using Euclidean distance formula and is as specified in the equation below:

Region marking is done on updated image. A boundary rectangle is drawn on each dominant region selected. The area of boundary rectangle is used in determining normalized area of dominant region. Then the location of the region is determined. Image path, number of regions present, each regions’ information like color, normalized area and location are stored in a meta-file for further processing. This file information is used for constructing Image index tree. When the search engine is initiated, index tree is constructed.

Figure 1:Illustration of assign-color.

Figure 2:Image with Boundaries marked.

III.Proposed method

We saw a statistical approach to measure the performance of system. We defined it in the terms of our problem, i.e. feature segmentation. We noticed that the success estimation is not singular, and has to be chosen by the user. This is due to the tradeoff between the PD and the FAR. In addition, we saw that sometimes the PD and FAR alone cannot give us enough information on the success estimation. If the amount of hand-marked data is small, the results need to be treated with care. We presented a way to normalize these results, which will give more weight to data that was marked extensively than to data that was marked sparsely. We are now ready to examine several parameters that affect the success of the system. We will be able to compare their performances and choose the ones that best suit our needs. The program method is based on the differences between urban color and natural color images. We compute several features color, shape based on gray level spatial dependencies. The features constitute a multidimensional space in which we seek a separator between the urban and nature image features. There are several parameters that may influence the success estimation of the program. Choosing different features, defining different thresholds, and other fixed parameters may yield different results. We tried to find the best combination of features and parameters, dedicated to urban and non-urban type of image features that gives the best segmentation.

The paper database’s definition was: Aerial images that contain urban and non-urban areas. Since Recognition’s database is classified and we were not able to gain access to it, we had to build our own database and find images that fit their demands.

The first step was to understand the meaning of the term "Urban". The definition that we got from Recognition was that "urban areas" mean "man-made" areas like buildings, roads etc., while "nature" is all the rest (trees, sky, sand etc.). It is important to understand that because of the way the program works, this question of defining urban areas becomes no more then a technical question, since the computer "learns" what is defined as "urban" using an example that we feed it, and according to this example the computer classifies a new input image. The computer detects an urban area if it has similar characteristics as the example that we gave it. For this reason, each pattern type that we want the program to detect as urban, must be given to it as an example during its learning process.

The system is trained by different types of images and features were stored in a clustered format using a clustering algorithm as shown below.

Clustering is the task of grouping the objects of a database into meaningful subclasses (that is, clusters) so that the members of a cluster are as similar as possible whereas the members of different clusters differ as much as possible from each other. Applications of clustering in spatial databases are, e.g., the detection of seismic faults by grouping the entries of an earthquake catalog or the creation of thematic maps in geographic information systems by clustering feature vectors. We can support clustering algorithms by our database primitives if the clustering algorithm is based on a "local" cluster condition, i.e. if it constructs clusters by analyzing a restricted neighborhood of the objects.

Data partitioning algorithms, which divide data into several subsets. Because checking all possible subset systems is computationally infeasible, certain greedy heuristics are used in the form of iterative optimization. Specifically, this means different relocation schemes that iteratively reassign points between the k clusters. Unlike traditional hierarchical methods, in which clusters are not revisited after being constructed, relocation algorithms gradually improve clusters. With appropriate data, this results in high quality clusters.

One approach to data partitioning is to take a conceptual point of view that identifies the cluster with a certain model whose unknown parameters have to be found. More specifically, probabilistic models assume that the data comes from a mixture of several populations whose distributions and priors we want to find. One clear advantage of probabilistic methods is the interpretability of the constructed clusters. Having concise cluster representation also allows inexpensive computation of intra-clusters measures of fit that give rise to a global objective function.

IV.System modeling

Figure 2: System block diagram

The images were taken from the web site of "Peace Now". This database is used for two reasons:

1) The images contain mostly wide natural areas with spotted "settled parts", which fit Recognition requirements.

2) The images contain different types of landscapes and building structures.

The images in the database are divided into two sets:

1) Training set – these images are manually marked and are given to the computer as an example for urban and non-urban areas.

2) Testing set – images that the computer automatically marks according to calculations done using the training group.

The used database is as shown below;



[1] V.N.Gudivada and V.V.Raghavan.: Special issue on content-based image retrieval systems - guest eds. IEEE Computer. 28(9) (1995) 18-22

[2] M.De Marsicoi, L.Cinque, and S.Levialdi.: Indexing pictorial documents by their content: a survey of current techniques. Image and Vision Computing. 15(2) (1997) 119-141.

[3] R.C.Gonzalez and P.Wintz.:Digital Image Processing. 2nd Edition.Addison- Wesley. Reading. Mass. (1987)

[4] M.Flickner etal.,: Query by image and video content: the QBIC system. IEEE Computer. 28(9) (1995) 23-32.

[5] D.Mohamad, G.Sulong and S.S.Ipson.: Trademark Matching using Invariant Moments. Second Asian Conference on Computer Vision. 5-8 Dec, Singapore. (1995).

[6] Cortelazzo etal.,:Trademark Shapes Description by String-Matching Techniques. Pattern Recognition. 27(8) (1994) 1005-1018.

[7] A.K.Jain and A.Vailaya.:Image Retrieval using Color and Shape. Second Asian Conference on Computer Vision. 5-8 Dec.Singapore. (1995) 529-533

[8] R.Mehrotra and J.E.Gary.: Similar-Shape Retrieval in Shape Data Management. IEEE Computer. 28(9) (1995)

[9] H.V.Jagadish.: A Retrieval Technique for Similar Shapes. Proceedings of ACM SIGMOD. Colorado. ACM. New York. May (1991) 208-217

[10] G.Lu and A.Sajjanhar.: Region-Based Shape Representation and Similarity Measure Suitable for Content-Based Image Retrieval. Multimedia Systems. (7) (1999) 165-174

[11] K.C.Ravishankar, B.G.Prasad, S.K.Gupta and K.K.Biswas.: Dominant Color Region Based Indexing Technique for CBIR. Proceedings of the International Conference on Image Analysis and Processing (ICIAP'99). Venice. Italy. Sept. (1999) 887-892