A Visual Attention Model Based on DCT Domain
Yiwei Jiang, De Xu · 2005
In the last few years, computer vision research has had an increased interest in modelling visual attention and a number of computable models of attention have been developed. But most of these models are based on pixel domain. Compressed domain technique deals with data directly in compressed domain. Computational complexity is greatly reduced owing to avoiding the expensive inverse DCT computation required to convert values from frequency domain to the pixel domain. In this paper, we present a method of detecting regions of interests (ROIS) in DCT domain. We use the properties of DCT coefficients to get low-level features of frames and combining feature maps into a unique saliency map. We use MPEG4 (Xvid) video clips to test our algorithm. The experiment results show that our algorithm can detect ROIS from DCT domain successfully.