Optimization of bottom-up saliency detection through gram polynomial decimation

摘要

Conspicuous objects in a particular scene grab human attention which subsequently avoid surroundings from being processed in detail. This hypothesis is modelled as saliency detection algorithms which have found vast range of applications from object tracking to data compression. This paper presents an optimization technique for bottom-up saliency detection algorithm based on Gram polynomial decimation. Gram polynomial basis offer more effective decimation of images as compared to Fourier basis by suppressing Gibbs error. The technique has been validated on MSRA10k Salient Object Database using Itti's method, graph based method, spectral residual approach and frequency-tuned method. Results show significant performance boost in terms of higher F₁ score and reduced computation time.