Human-inspired shape-based image retrieval using weighted City-block distance and Fourier descriptors
Contour-based Fourier descriptors are very simple and effective shape description method used for content-based image retrieval. Similarity between Fourier descriptors is usually computed using measures such as City-block or Euclidean distance. These similarity measures consider all harmonics to be equally important, therefore harmonics with larger magnitude tend to have larger significance during the computation of shape similarity. In order to increase the importance of harmonics with lower magnitude, we propose to use weighted City-block distance for computing shape similarity. The proposed weighting coefficients are inspired by the contrast sensitivity of the human visual system to different spatial frequencies, known as the Contrast Sensitivity Function (CSF). Although weighted distances generally do not improve the retrieval performance, experimental results clearly demonstrate that human observers favour the retrieval system based on the weighted distances, and find it more accurate and relevant.