Study On Visual Saliency Detection In Compressed Domain And Its Applications

Full decoding is required to be performed beforehand by using the existing image/video saliency detection algorithms based on pixel domain. This will lead to high computational complexity. Nowadays most of the images/videos are encoded by using discrete cosine transform (DCT) technique. Using the compressed data to realize visual saliency detection, the saliency detection algorithms based on compressed domain can significantly decreases the computational complexity and improves the applications in mobile terminals confined by limited resources. Therefore, it is worth further studying visual saliency detection based on compressed domain. This paper focuses on the image resizing based on the saliency detection in compressed domain, as well as the motion saliency detection in compressed domain. The results obtained are as follows.This paper proposes an image resizing algorithm based on visual saliency detection in DCT domain. In the presented method, a saliency detection model in DCT domain is utilized to obtain the saliency map. Then the saliency map and the energy map are employed to implement seam carving (SC). The experiments are conducted in the public image database. And the experimental results show that the proposed algorithm can not only protect the important contents, but also guarantee the integrity of visual contents. And the average quality index of the proposed algorithm is higher 8.99% than that of the existing algorithm.And this paper proposes a motion saliency detection algorithm based on compressed domain. The motion vectors for each 8x8 DCT block in prediction frames are obtained from the compressed video bit stream, which are used to extract motion feature. The motion magnitude map is calculated by using the motion vector value. And the motion center surround map is calculated by the anisotropic Gaussian distribution function. Then the two maps are fused to a motion saliency map with uncertain weights. The motion saliency map may be combined with the static saliency map to achieve the final saliency map. The area under the receiver operating characteristic curve (AUC) and F-measure are utilized to evaluate the performance of the proposed algorithm by using standard test sequence databases. The experimental results show that the subjective effect of the proposed motion saliency map and fusion saliency map are more significant. Compared with the existing algorithms, the average of AUC and F-measure values by using the presented method are higher 11.7% and 11.4%, respectively. And the average of AUC and F-measure values of the corresponding fusion saliency map are higher 4.3% and 25.7%, respectively.