Shape matching and object recognition using chord contexts

We propose a new effective shape descriptor, chord context, for shape description in content-based image retrieval. For a shape, the chord context describes a frequency distribution of chord lengths with different orientations. The histograms which represent the chord context are compacted and normalized into a feature matrix. Unlike other shape representation schemes, the proposed scheme is able to extract attributes from both contour as well as region information without the need for special landmarks or key-points. It does not require points on the edge with their orders, but it can capture the feature of shapes with holes or even with separated regions. In addition, the proposed method is shown to be unaffected by image translation, rotation and scaling; at the same time, it is robust to minor occultation, non-rigid deformations, distortions and corruption due to noise. Several experimental results demonstrate the feasibility of the chord context descriptor methodology and also highlight its advantages over other existing methodologies.