Project topics and materials
Get Complete Project Material File(s) Now! »
Image content retrieval and mining overview
CBIR technology has been used in several applications, for example ngerprint identi- cation, biodiversity information systems, digital libraries, medicine, historical research, etc. The creation of CBIR systems involves research on databases and image processing, storage handling problems and user friendly interfaces. In these systems, image processing algorithms are used to extract features that represent the image properties such as color, texture and shape. The CBIR approach retrieves images similar to one chosen by the user (query-by-example) based on a similarity metric. Images are particularly complex to manage besides the volume they occupy: retrieval of images from databases is a context-dependent task. The rst step towards retrieval requires a translation of high-level user perception into low-level image features. Moreover, image indexing is not just a string processing method: images or objects in images are represented as points in a high-dimensional space. CBIR systems work have two main functionalities: data insertion and query processing. The data insertion system is responsible for extracting features from the images and storing them in a feature database. The query processing module is organized with an interface that allows the user to specify a query image and visualize the retrieved similar images. The query processing module extracts feature vectors from the query image and applies a metric (such as the L2 or the L1 norm) to evaluate the similarity between the query and the database images. The database images are ranked in decreasing order of similarity to the query image and the most similar images are shown in the output interface.
The IKONA system
The IMEDIA research group at INRIA Rocquencourt [INRI] has developed a contentbased indexing technique and interactive retrieval methods for browsing large multimedia databases by content. In order to design an eective image retrieval system, the database is divided into two categories. The rst category is concerned with speci c databases with known ground truth. During the process of indexing the user will consider these ground truths in order to tune the models or the parameters which in turn maximize the system eciency. The group has developed specic signatures for face recognition and ngerprint identication. The second category is concerned with databases containing heterogeneous images without any ground truth. In this context, generic image features are computed which describe a general visual appearance such as color and texture [Bouj 01, Goue 01].
The IKONA system architecture is based on a client-server architecture where the server is written in C++ and includes image feature extraction algorithms, user interaction (retrieval with visual similarity, relevance feedback mode, partial query mode, points of interest mode, etc…) and a network module for communication with the client. The client needs to be portable and is written in Java. It runs on machines with the Java Runtime Environment (JRE). The user is presented with a Graphical User Interface (GUI), the query mode is set for the server and the relevant images are displayed as the search results. The system is exible in the sense that functionalities can be easily added without disturbing the overall architecture. In order to deal with generic databases, IKONA includes a relevance feedback technique which allows the user to rene their query by specifying a set of relevant and non-relevant images IKONA has a Relevance Feedback (RF) mode [Fere 05b, Fere 05a] for category search in image databases which assists the user in rapidly nding images in a large database. The IKONA system also has a region based query mode, in which the user can select a part of the image and the system will search images or part of images that are visually similar to the selected part [Fauq 02, Fauq 04, Fere 04]. In this case the query is focussed and the system response is enhanced with regard to the user’s objective since background image properties are not considered. Several segmentation methods and point of interest methods have been developed to achieve partial queries. The generic image database with query image selection and query results is shown in Figure 2.1.
Texture content of an image
Texture is an implicit property of all surfaces, natural or man-made. For instance, clouds have a dierent textural property to roads. Texture also provides us with the structural arrangement of objects and their relationships with the surroundings. A vast amount of research has been done in this eld over the past few decades. The co-occurrence matrix representation of texture [Hara 73] three decades ago opened a new direction in the eld of pattern recognition. This approach considers texture as grey level spatial dependency. The matrix is constructed by taking the distance
between image pixels. Various statistics are then computed from this matrix in order to represent texture properties. Some of these statistics lack the essence of human visual perception of texture. Motivated by this fact a new texture representation was developed by [Tamu 78]. These features are closer to human perception of texture.
Table of contents :
Contents
List of gures
List of tables
1 Introduction
1.1 Image content retrieval and mining overview
1.2 Which information ?
1.3 The methodology
1.4 Organization of the thesis
2 Image Information Mining Systems
2.1 Systems
2.2 The IKONA system
2.3 The KIM system
2.4 Image representation and similarity
2.5 Image content
2.5.1 Shape content of an image
2.5.2 Texture content of an image
2.6 Image classication
2.6.1 Classication tools
2.7 Discussion
3 Road Network Structural Information
3.1 Introduction
3.2 Network extraction and representation
3.3 Preliminary study
3.3.1 Feature selection and clustering
3.4 Larger image database with rened classes and features
3.4.1 New features from the graph
3.4.2 Classication
3.5 Discussion
4 Region Structural Information
4.1 Urban region structures
4.2 Segmentation of urban regions
4.2.1 Features from the urban regions
4.2.2 Classication
4.2.3 2-Level classication
4.2.4 Classication and retrieval with relevance feedback
4.3 Discussion
5 Dependence of primitive features on image resolution and size
5.1 Introduction
5.2 Road network and urban region extraction for dierent resolutions
5.3 Classication
5.4 Discussion
6 Indexing of large satellite images
6.1 Introduction
6.2 Indexing
6.2.1 Step 1: The database
6.2.2 Step 2: The feature le
6.2.3 Step 3: The classication
6.2.4 Detailed analysis of 4 regions
6.3 Indexing results on large satellite images
6.3.1 Barcelona
6.3.2 Detailed analysis of 2 regions
6.3.3 Los Angeles
6.3.4 Detailed analysis of 2 regions
6.3.5 Madrid
6.3.6 Detailed analysis of 2 regions
6.3.7 Paris
6.3.8 Detailed analysis of 2 regions
6.4 Discussion
7 Conclusion
A Personal Biography
Bibliography
