image processing3 Presentation Seminar

[born_for seminars]

ABSTRACT

Retrieval efficiency and accuracy are two important issues in designing a content-based data retrieval system. The need for developing shape feature that are better able to capture human perceptual similarity of shapes are required. Developing an automatic retrieval algorithm, which matches human performance, is an extremely difficult and challenging task. However considering the substantial amount of time and effort needed for a manual retrieval from a large image database an automatic shape based retrieval technique can significantly simplify the retrieval task. We propose a method for trademark image database retrieval based on object shape information that would supplement traditional text based retrieval system. The proposed system achieves both the desired efficiency and accuracy using a two stage hierarchy: in the first stage simple and easily computable shape features are used to quickly browse through the database to generate a moderate number of possible retrievals when a query is presented and in the second stage, images from the first stage are screened using a deformable template matching process to discard spurious matches.


Keywords: Image database, trademarks, deformable temaplate, moment invarients, shape similarity, Retrievation.




I. INTRODUCTION

Information is inherently multimodel. Humans can efficiently and effectively process information simultaneously in multiple dimensions. These multiple medias that aid effective communication, can be characterized into speech, audio, image, video, and textual data.
Digital images are convenient media for describing and storing spatial, temporal, spectral, and physical components of information contained in variety of domains. A typical database consists of hundreds of thousands of images, taking up gigabytes of memory space. While advances in image compression algorithms have alleviated the storage requirement to some extent, large volumes of these images make it difficult for a user to quickly browse through the entire database. Therefore, an efficient and automatic procedure is required for indexing and retrieving images from databases.
Traditionally, textual features such as filenames, captions, and keywords have been used to annotate and retrieve images. But there are several problems with these methods.

LITERATURE REVIEW:

Much of the past research in CBIR has concentrated on the feature extraction stage. For a given query image, its feature vector is calculated and those images which are most similar to query based on appropriate distance measure in the feature space are retrieved. But these traditional schemes are in general too expensive and not meaningful for such an application. Various schemes have been proposed for shape representation and retrieval. These include shape representation using polygonal approximation of shape [1] and matching using the polygonal vertices, image representation on the basis of strings [2] [3], comparing the images using the hausdorff distance [4].

The above techniques relay on a single model and its associated features to describe the object shape. Many of the shape features are not invariant to large variations in image size, position, and orientation. When we consider large image database, retrieval speed is important consideration. We, therefore, need to identify shape features which can be efficiently computed and which are invariant to 2D rigid transformations.

Approach:

In this paper we address the problem of efficiently and accurately retrieving images from a database of trademark images purely based on shape analysis. Since we desire a system that has both high speed and accuracy of retrievals, we propose a two-tired hierarchical image retrieval system.
Figure 1 shows a block diagram of our proposed retrieval system.





The first stage computes simple image features from input query image to prune the database to a reduced set of plausible matches. As simple shape features are used in screening the image database, this first stage can be performed very fast; this output is then presented to a detailed matcher in the second stage. This stage uses a deformable template model to eliminate false matches. Our retrieval system is insensitive to variations in scale, rotation, and translation.

2. TRADEMARKS:

Trademarks represent a gamut of pictorial data. There are over a million registered trademarks. A trademark is a word, phrase, symbol or design, or combination of words, phrases, symbols or designs, which identifies and distinguishes the source of goods or services of one party from those of others. Most of the trademarks are an abstract representation of a concept in a world, like an abstract drawing of an animal, or natural objects (sun, moon, etc.) [5]. It is extremely challenging and instructive to study and address the issue of image database retrieval on this huge source of complex pictorial data.

3. SHAPE BASED MATCHING:

The goal of proposed retrieval system is to present to the user a subset of database images that are visually similar to the query. A major challenge is that trademarks that appear to be perceptually similar need not be exactly similar in their shape. In fact, it is extremely difficult to define and capture perceptual similarity. It is extremely difficult for a computer vision system to identify objects in the image automatically. Consider for example two perceptually similar images (a) and (b).

4. HIERARCHICAL SYSTEM FOR EFFICIENT RETRIEVAL:

We describe a system to extract visually similar trademarks from a database of design marks. The goal of the system is to present to the user all possible similar design marks that resemble in shape to the query trademark. Our aim is to build a system with both high speed and accuracy; we use a hierarchical two-level feature extraction and matching structure for image retrieval (Fig. 1). Our system uses multiple shape features for initial pruning stage. Retrievals based on these features are integrated for better accuracy [6] .The second stage uses deformable template matching to eliminate false retrievals present among the output of first stage, there by improving the precision rate of the system.

4.1 IMAGE ATTRIBUTES:

In order to retrieve images, we must be able to efficiently compare two images to determine if they have a similar content.
Let {F(x,y); x,y = {1,2,……..,N} be a two dimensional pixel array.
Let f represent a mapping from the image space onto the n-dimensional Feature space
f = {x1,x2,…….,xn}, i.e.,

where n is the number of features used to represent the image .The difference between two images,F1 and F2,cab be expressed as the distance,D,between the respective feature vectors,x1 and x2.The choice of this distance measure ,D is critical and domain-dependent. The problem of retrieval can then be posed as follows: Given a query image P, retrieve a subset of the images, from the image database, ,such that

Where t is user specified threshold. Alternatively, instead of specifying the threshold, a user can ask the system to output, say the top-ten images, which are most similar to the query image.

4.2 FAST PRUNING STAGE:

It is desirable to have an image retrieval system which is insensitive to large variations in image scale, rotation, and translation. Hence, the pruning stage has to be not only fast but should also extract invariant features for matching. The following shape features are used

Edge Variation:

A histogram of the edge directions is used to describe global shape information.

Invariant Moments:

The global image shape is described in terms of seven invariant moments.