Chord Context Algorithm for Shape Feature Extraction [chapter]

Yang Mingqiang, Kpalma Kidiyo, Ronsin Joseph
2011 Object Recognition  
Introduction The emergence of new technologies makes it easy to generate information in visual forms, leading everyday to an increasing number of generated digital images. At the same time, the rapid advances in imaging technologies and the widespread availability of Internet access motivate data browsing into these data bases. For image description and retrieval, manual annotation of these images becomes impractical and inefficient. Image retrieval is based on observation of an ordering of
more » ... h scores obtained by searching through a database. The key challenges in building a retrieval system are the choice of attributes, their representations, query specification methods, match metrics and indexing strategies. A large number of retrieval methods using shape descriptors has been described in literature. Compared to other features, for example, color or texture, object shape is unique. It enables us to recognize an object without further information. However, since shapes are 2D images that are projections of 3D objects, the silhouettes may change from one viewpoint to another with respect to objects and non-rigid object motion (e.g., walking people or flying bird) and segmentation errors caused by lighting variations, partial occultation, scaling, boundary distortion and corruption by noise are unavoidable. As we know, while computers can easily distinguish slight differences between similar objects, it is very difficult to estimate the similarity between two objects as perceived by human beings, even when considering very simple objects. This is because human perception is not a mere interpretation of a retinal patch, but an active interaction between the retinal patch and a representation of our knowledge about objects. Thus the problem is complicated by the fact that a shape does not have a mathematical definition that exactly matches what the user feels as a shape. Solutions proposed in the literature use various approaches and emphasize different aspects of the problem. The choice of a particular representation scheme is usually driven by the need to cope with requirements such as robustness against noise, stability with respect to minor distortions, and invariance to common geometrical transforms or tolerance to occultation, etc. For general shape representation, a recent review is given in [1] [2] . In this chapter, a shape descriptor based on chord context is proposed. The basic idea of chord context is to observe the lengths of all parallel and equidistant chords in a shape, and Object Recognition 66 to build their histogram in each direction. The sequence of vector features extracted forms the feature matrix for a shape descriptor. Because all the viewpoint directions, considered with a certain angle interval, are chosen to produce the chord length histogram, this representation is unlike conventional shape representation schemes, where a shape descriptor has to correspond to key points such as maxima of curvature or inflection points, for example, Smooth Curve Decomposition [3], Convex Hull [4], Triangle-area representation (TAR) [5] and Curvature Scale Space (CSS) [6] [7] etc. The proposed method needs no special landmarks or key points. There is also no need for certain axes of a shape. The proposed descriptor scheme is able to capture the internal details, specifically holes, in addition to capturing the external boundary details. A similarity measure is defined over chord context according to its characteristics and it confirms efficiency for shape retrieval from a database. This method is shown to be invariant under image transformations, rotations, scaling and robust to non-rigid deformations, occultation and boundary perturbations by noise thus it is well-adapted to shape description and retrieval. In addition, the size of the descriptor attribute is not very great; it has low-computational complexity compared to other similar methods.
doi:10.5772/14979 fatcat:vlfyi4yrhvgvvpzvqxpklscxgu