Other Free Encyclopedias » Online Encyclopedia » Encyclopedia - Featured Articles » Contributed Topics from F-J

Image Retrieval - Existing Techniques, Content-Based (CBIR) Systems

features visual data color

William I. Grosky
Department of Computer and Information Science
University of Michigan-Dearborn, Dearborn, MI, USA

Definition : Image retrieval techniques integrate both low-level visual features, addressing the more detailed perceptual aspects, and high-level semantic features underlying the more general conceptual aspects of visual data.

The emergence of multimedia technology and the rapid growth in the number and type of multimedia assets controlled by public and private entities, as well as the expanding range of image and video documents appearing on the web, have attracted significant research efforts in providing tools for effective retrieval and management of visual data. Image retrieval is based on the availability of a representation scheme of image content. Image content descriptors may be visual features such as color, texture, shape, and spatial relationships, or semantic primitives.

Conventional information retrieval is based solely on text, and these approaches to textual information retrieval have been transplanted into image retrieval in a variety of ways, including the representation of an image as a vector of feature values. However, “a picture is worth a thousand words.” Image contents are much more versatile compared with text, and the amount of visual data is already enormous and still expanding very rapidly. Hoping to cope with these special characteristics of visual data, content-based image retrieval methods have been introduced. It has been widely recognized that the family of image retrieval techniques should become an integration of both low-level visual features, addressing the more detailed perceptual aspects, and high-level semantic features underlying the more general conceptual aspects of visual data. Neither of these two types of features is sufficient to retrieve or manage visual data in an effective or efficient way. Although efforts have been devoted to combining these two aspects of visual data, the gap between them is still a huge barrier in front of researchers. Intuitive and heuristic approaches do not provide us with satisfactory performance. Therefore, there is an urgent need of finding and managing the latent correlation between low-level features and high-level concepts. How to bridge this gap between visual features and semantic features has been a major challenge in this research field.

The different types of information that are normally associated with images are:

  • Content-independent metadata: data that is not directly concerned with image content, but related to it. Examples are image format, author’s name, date, and location.
  • Content-based metadata:

    • Non-information-bearing metadata: data referring to low-level or intermediate-level features, such as color, texture, shape, spatial relationships, and their various combinations. This information can easily be computed from the raw data.
    • Information-bearing metadata: data referring to content semantics, concerned with relationships of image entities to real-world entities. This type of information, such as that a particular building appearing in an image is the Empire State Building , cannot usually be derived from the raw data, and must then be supplied by other means, perhaps by inheriting this semantic label from another image, where a similar-appearing building has already been identified.

Low-level visual features such as color, texture, shape and spatial relationships are directly related to perceptual aspects of image content. Since it is usually easy to extract and represent these features and fairly convenient to design similarity measures by using the statistical properties of these features, a variety of content-based image retrieval techniques have been proposed. High-level concepts, however, are not extracted directly from visual contents, but they represent the relatively more important meanings of objects and scenes in the images that are perceived by human beings. These conceptual aspects are more closely related to users’ preferences and subjectivity. Concepts may vary significantly in different circumstances. Subtle changes in the semantics may lead to dramatic conceptual differences. Needless to say, it is a very challenging task to extract and manage meaningful semantics and to make use of them to achieve more intelligent and user-friendly retrieval.

High-level conceptual information is normally represented by using text descriptors. Traditional indexing for image retrieval is text-based. In certain content-based retrieval techniques, text descriptors are also used to model perceptual aspects. However, the inadequacy of text description is very obvious:

  • It is difficult for text to capture the perceptual saliency of visual features.
  • It is rather difficult to characterize certain entities, attributes, roles or events by means of text only.
  • Text is not well suited for modeling the correlation between perceptual and conceptual features.
  • Text descriptions reflect the subjectivity of the annotator and the annotation process is prone to be inconsistent, incomplete, ambiguous, and very difficult to be automated.

Although it is an obvious fact that image contents are much more complicated than textual data stored in traditional databases, there is an even greater demand for retrieval and management tools for visual data, since visual information is a more capable medium of conveying ideas and is more closely related to human perception of the real world. Image retrieval techniques should provide support for user queries in an effective and efficient way, just as conventional information retrieval does for textual retrieval. In general, image retrieval can be categorized into the following types:

  • Exact Matching – This category is applicable only to static environments or environments in which features of the images do not evolve over an extended period of time. Databases containing industrial and architectural drawings or electronics schematics are examples of such environments.
  • Low-Level Similarity-Based Searching – In most cases, it is difficult to determine which images best satisfy the query. Different users may have different needs and wants. Even the same user may have different preferences under different circumstances. Thus, it is desirable to return the top several similar images based on the similarity measure, so as to give users a good sampling. The similarity measure is generally based on simple feature matching and it is quite common for the user to interact with the system so as to indicate to it the quality of each of the returned matches, which helps the system adapt to the users’ preferences. Figure 1 shows three images which a particular user may find similar to each other. In general, this problem has been well-studied for many years.
  • High-Level Semantic-Based Searching – In this case, the notion of similarity is not based on simple feature matching and usually results from extended user interaction with the system. Figure 2 shows two images whose low-level features are quite different, yet could be semantically similar to a particular user as examples of peaceful scenes. Research in this area is quite active, yet still in its infancy. Many important breakthroughs are yet to be made.

For either type of retrieval, the dynamic and versatile characteristics of image content require expensive computations and sophisticated methodologies in the areas of computer vision, image processing, data visualization, indexing, and similarity measurement. In order to manage image data effectively and efficiently, many schemes for data modeling and image representation have been proposed. Typically, each of these schemes builds a symbolic image for each given physical image to provide logical and physical data independence. Symbolic images are then used in conjunction with various index structures as proxies for image comparisons to reduce the searching scope. The high-dimensional visual data is usually reduced into a lower-dimensional subspace so that it is easier to index and manage the visual contents. Once the similarity measure has been determined, indexes of corresponding images are located in the image space and those images are retrieved from the database. Due to the lack of any unified framework for image representation and retrieval, certain methods may perform better than others under differing query situations. Therefore, these schemes and retrieval techniques have to be somehow integrated and adjusted on the fly to facilitate effective and efficient image data management.

Existing Techniques

Visual feature extraction is the basis of any content-based image retrieval technique. Widely used features include color, texture, shape and spatial relationships. Because of the subjectivity of perception and the complex composition of visual data, there does not exist a single best representation for any given visual feature. Multiple approaches have been introduced for each of these visual features and each of them characterizes the feature from a different perspective.

Color is one of the most widely used visual features in content-based image retrieval. It is relatively robust and simple to represent. Various studies of color perception and color spaces have been proposed, in order to find color-based techniques that are more closely aligned with the ways that humans perceive color. The color histogram has been the most commonly used representation technique, statistically describing combined probabilistic properties of the various color channels (such as the®ed, (G)reen, and (B)lue channels), by capturing the number of pixels having particular properties. For example, a color histogram might describe the number of pixels of each red channel value in the range [0, 255]. Figure 3 shows an image and three of its derived color histograms, where the particular channel values are shown along the x-axis, the numbers of pixels are shown along the y-axis, and the particular color channel used is indicated in each histogram. It is well known that histograms lose information related to the spatial distribution of colors and that two very different images can have very similar histograms. There has been much work done in extending histograms to capture such spatial information. Two of the well-known approaches for this are correlograms and anglograms. Correlograms capture the distribution of colors of pixels in particular areas around pixels of particular colors, while anglograms capture a particular signature of the spatial arrangement of areas (single pixels or blocks of pixels) having common properties, such as similar colors. We note that anglograms also can be used for texture and shape features.

Texture refers to the patterns in an image that present the properties of homogeneity that do not result from the presence of a single color or intensity value. It is a powerful discriminating feature, present almost everywhere in nature. However, it is almost impossible to describe texture in words, because it is virtually a statistical and structural

property. There are three major categories of texture-based techniques, namely, probabilistic/statistical, spectral , and structural approaches. Probabilistic methods treat texture patterns as samples of certain random fields and extract texture features from these properties. Spectral approaches involve the sub-band decomposition of images into different channels, and the analysis of spatial frequency content in each of these sub-bands in order to extract texture features. Structural techniques model texture features based on heuristic rules of spatial placements of primitive image elements that attempt to mimic human perception of textural patterns.

The well known Tamura features include coarseness, contrast, directionality, line-likeness, regularity, and roughness. Different researchers have selected different subsets of these heuristic descriptors. It is believed that the combination of contrast, coarseness, and directionality best represents the textural patterns of color images. Figure 4 illustrates various textures.

Shape representation is normally required to be invariant to translation, rotation, and scaling . In general, shape representations can be categorized as either boundary-based or region-based . A boundary-based representation uses only the outer boundary characteristics of the entities, while a region-based representation uses the entire region. Shape features may also be local or global . A shape feature is local if it is derived from some proper subpart of an object, while it is global if it is derived from the entire object. See Figure 5 for an illustration of these concepts.

A combination of the above features are extracted from each image and transformed into a point of a high-dimensional vector space. Using this representation, the many techniques developed by the information retrieval community can be used to advantage. As the dimensionality of the underlying space is still quite high, however, the many disadvantages caused by the curse of dimensionality also prevail.

Originally devised in the context of estimating probability density functions in high-dimensional spaces, the curse of dimensionality expresses itself in high-dimensional indexing by causing log time complexity indexing approaches to behave no better than linear search as the dimensionality of the search space increases. This is why there has been so much effort spent in the development of efficient high-dimensional indexing techniques, on the one hand, and in dimensional reduction techniques which capture the salient semantics, on the other hand.

As the ultimate goal of image retrieval is to serve the needs and wants of users who may not even know what they are looking for but can recognize it when they see it, there has been much work done in trying to discover what is in the mind of the user. A very common technique for this is relevance feedback . Originally advanced in the information retrieval community, it has become a standard in most existing image retrieval systems, although some researchers believe that more involved user interactions are necessary to discover user semantics. This technique helps the system refine its search by asking the user to rank the returned results as to relevance. Based on these results, the system learns how to retrieve results more in line with what the user wants. There have been many new approaches developed in recent years, but the classical techniques are query refinement or feature reweighting . Query refinement transforms the query so that more of the positive and less of the negative examples will be retrieved. Feature reweighting puts more weight on features which help to retrieve positive examples and less weight on features which aid in retrieving negative examples. This process continues for as many rounds as is necessary to produce results acceptable to the user.

Needless to say, human beings are much better than computers at extracting and making use of semantic information from images. Many researchers believe that complete image understanding should start from interpreting image objects and their relationships. The process of grouping low-level image features into meaningful image objects and then automatically attaching correlated semantic descriptions to image objects is still a challenging problem in image retrieval. One of the earliest examples of such an approach is that used in the ImageMiner system. Their method is structural in nature, using graph grammars, and generates scene descriptions with region labels. Current techniques in this area use Baysian approaches which integrate textual annotations and image features.

Content-Based Image Retrieval (CBIR) Systems

There are several excellent surveys of content-based image retrieval systems. We mention here some of the more notable systems. The first, QBIC (Query-by-Image-Content), was one of the first prototype systems. It was developed at the IBM Almaden Research Center and is currently folded into DB2. It allows queries by color, texture, and shape, and introduced a sophisticated similarity function. As this similarity function has a quadratic time-complexity, the notion of dimensional reduction was discussed in order to reduce the computation time. Another notable property of QBIC was its use of multidimensional indexing to speed-up searches. The Chabot system, developed at the University of California at Berkeley, brought text and images together into the search task, allowed the user to define concepts, such as that of a sunset , in terms of various feature values, and used the post-relational database management system Postgres. Finally, the MARS system, developed at the University of Illinois at Urbana-Champaign, allowed for sophisticated relevance feedback from the user.

Image Search Engine - Abstract, Introduction, Web, Collection-based search engine, Content-based image retrieval, Conclusion [next] [back] Image Manipulation: Science Fact or Fiction - Suggested General Guidelines for Scientific Imaging Applications

User Comments

Your email address will be altered so spam harvesting bots can't read it easily.
Hide my email completely instead?

Cancel or

Vote down Vote up

over 5 years ago


Vote down Vote up

almost 6 years ago