Rough Feature Selection for CBIR. Outline

Transcription

1 Rough Feature Selection for CBIR Instructor:Dr. Wojciech Ziarko presenter :Aifen Ye 19th Nov., 2008 Outline Motivation Rough Feature Selection Image Retrieval Image Retrieval with Rough Feature Selection Summary 1

2 Motivation semantic gap between user and machine CBIR mainly use only three kind of features-color, shape, texture, but the results of retrieval are not satisfying. Features commonly used do not fully represent the visual properties of image Need to select other kind of features to apply to CBIR Rough feature selection is promising and effective Rough Feature Selection Rough set theory has been introduced by Pawlak to deal with imprecise or vague concepts. rapid growth of interest in rough set theory and its applications. 2

3 Rough Feature Selection Rough Sets Theory is a mathematical tool that had been used successfully to discover data dependencies and reduce the number of attributes Reducts that are obtained by using Rough Sets are very informative and all the other attributes can be removed with a minimal information loss due to the use of the degree of dependency measure Rough Feature Selection Feature selection process refers to choosing subset of attributes from the set of original attributes. The purpose of the feature selection is to identify the significant features, eliminate the dispensable features and build good classification rules. 3

4 Rough Feature Selection The benefits of feature selection are twofold: considerably decreases the computation time increases the accuracy of the resulting mode Rough Feature Selection Let A = (U;A) be an information system Assuming P and Q are equivalence relations in U, the positive region POS P (Q) is defined as POS ( Q) = PX (1) X U / Q Q depends on P in a degree of,if P γ ( Q) = P POS ( Q) P U γ P ( Q)(0 γ ( Q) 1) P (2) 4

5 Rough Feature Selection A reduct is defined as a subset X of the conditional attribute set C such that γ ( D) = γ ( D) X (3) Where D is the decision attribute the set R of all reducts is defined as: R = C { X X C, γ ( D) = γ ( D) } X C (4) Rough Feature Selection For Rough Set Attribute Reduction, the minimal reduct Rmin R is defined as the set of any reduct searched in R with minimum cardinality: R min { X X R : Y R X Y } =, 5

6 Rough Feature Selection To select feature using rough set theory is to get the minimum subset of the entire conditional attributes, but maintain the same discrimination power. Rough Feature Selection Nowadays numerous successful implementations of feature selection using rough set theory are available. The Rough Feature Selection s application areas are very large, not only in the computer science area (such as Information retrieval), but also in lots of other studies( such as commercial application and biological analysis). The applications are well summarized by an Indian scholar K. Thangavel in Table 1 and 2 6

7 Image Retrieval Figure 1: view of the many facets of image retrieval as a field of research Image Retrieval What s image retrieval? An image retrieval system is a computer system for browsing, searching and retrieving images from a large database of digital images. Most traditional and common methods of image retrieval utilize some method of adding metadata such as captioning, keywords, or descriptions to the images so that retrieval can be performed over the annotation words. Manual image annotation is time-consuming, laborious and expensive; to address this, there has been a large amount of research done on automatic image annotation. content-based image retrieval (CBIR), which aims at avoiding the use of textual descriptions. Instead retrieves images based on their visual similarity to a user-supplied query image or user-specified image features 7

8 Category of Image Retrieval (IR) Semantic IR Content Based IR(CBIR) 15 Category of IR For the different ways to get description of image, the Semantic IR can be further divide into two categories: General Semantic IR Manually add description for each image in the database Searching and matching with the keywords, similar to the text information retrieval Auto-annotated semantic IR Images are annotated by certain algorithm developed for extracting annotation Perform as text information retrieval with the extracted annotation (example for semantic IR) 8

9 Category of IR Content Based IR(CBIR) Extract features to represent the visual properties, such as shape, color and texture, from the image itself. Searching digital images databases and Matching the extracted features to get the most similar images. Development of CBIR CBIR originated in 1992, when it was used by T. Kato to describe experiments into automatic retrieval of images from a database, based on the colors and shapes present. 9

10 Development of CBIR First commercial system-qbic system (Query By Image Content) from IBM color percentages, color layout, texture, shape, location, and keywords popular system- BlobWorld system Using segmentation to get regions as features--shape SIMBA system Using features invariant against rotation and translation, mainly concerning color and texture Web-based system-tineye Content based image Searching in internet. General CBIR Flow chart Object image Feature Extraction Feature matching Return and display images query image Images database Figure 2: General CBIR flow chart 10

11 General CBIR Figure 3: screenshot of CBIR system ImgSeek querying for flower image Figure 4: screenshot of CBIR system ImgSeek querying for sunset image 11

12 Figure 5: screenshot of CBIR system ImgSeek querying by draft image drew by user CBIR Potential uses for CBIR include: Art collections Photograph archives Retail catalogs Medical diagnosis Crime prevention The military Intellectual property Architectural and engineering design Geographical information and remote sensing systems 12

13 Features in IR Any kind of IR is a process of matching keywords (Semantic IR) or features (CBIR) to get the most similar image. Feature in CBIR The commonly used features in CBIR: Color Retrieving images based on color similarity is achieved by computing a color histogram for each image that identifies the proportion of pixels within an image holding specific values (that humans express as colors). Current research is attempting to segment color proportion by region and by spatial relationship among several color regions 13

14 Feature in CBIR Texture Texture measures look for visual patterns in images and how they are spatially defined. Texture is a difficult concept to represent. The identification of specific textures in an image is achieved primarily by modeling texture as a twodimensional gray level variation. Feature in CBIR shape Shape does not refer to the shape of an image but to the shape of a particular region that is being sought out. Shapes will often be determined first applying segmentation or edge detection to an image. 14

15 Features in CBIR Everything can be a feature, if it fulfils two conditions. Firstly, it should represent a visual property; secondly, it should be statistically independent to other features. Based on this view it is possible to argue for a large number of features to be reasonable. The feature problem is shifted from designing wellperforming features to estimating the relevance of a feature for a particular querying situation. Features in CBIR Features for CBIR can be divided into 12 groups: Basic Image pixel Features Color Histograms Invariant Features Invariant Features by Integration Invariant Feature Histograms Invariant Feature Vectors Invariant Fourier Mellin Features Gabor Features Tamura Features Global Texture Descriptor Local Features Histograms of Local Features Region-Based Features PCA Transformed Features Correlation of Different Features 15

16 Rough feature selection in CBIR Target image Feature Extraction Rough Feature selection Feature matching Return and display images query image Images database Figure 6: CBIR flow chart using rough feature selection Rough feature selection in CBIR Project target Extract a list of features from the original image Using Interactive CBIR to estimate the relevance of features Using rough feature selection to get the most efficient and simplest features for CBIR 16

17 Rough feature selection in CBIR Extract a list of features from the images Query from the image database and matching with the features Interactive select and denote relevance of each image Rough feature selection to select the most effective and simplest features Figure 7: further refinement of the purple rectangle area in Figure 6 Rough feature selection in CBIR E k -the k th returned image in the training process. A ik - the difference of the i th feature between the object image and the k th returned image D k -the relevance index of the object image and the k th returned image. It is manually denoted. E 1 E 2 E 3 Table 1: information table A 1 A 2 D 17

18 Rough feature selection in CBIR Step 1 Specify preference for each Attribute, the simpler, the more preferable. Step 2 Using rough set theory to get reduct of the decision table Step 3 Once the reduct is obtained, using the selected features to match and retrieve image. expectation more effective and simpler features for image retrieval Increase accuracy of CBIR retrieval result 18

19 summary Motivation of this project Introduce the rough feature selection Introduce the image retrieval, semantic IR and CBIR. Further discuss the features for image retrieval Give the design of Applying rough feature selection to CBIR Question? 19