Research Of Image Saliency Based On Wavelet Domain

Image saliency detection has a long history. It is a challenging research area because of imitation of human visual attention mechanism. Image saliency detection is to detect important image content and provide convenience for the image application. Now, image saliency algorithms have a series of problems, such as accuracy and complexity. Compared with the Fourier transform, Wavelet transform is partial transformation of the spatial and frequency which can extract image information efficiently and so can be used to detect image saliency.The main work accomplished as following:(1)By analyzing the Frequency-tuned salient region detection algorithm(FT), this paper proposes a novel method based on Contrast Sensitivity Function(CSF) in wavelet domain. CSF model may improve image contrast. So Combining CSF with wavelet analysis, this paper gets different frequency information and uses weighted Euclidean distance to calculate saliency value. Parabolic is used improves contrast finally.(2)Through the analysis of Bayesian saliency algorithm via low and middle level cues(BS), this paper proposes the saliency detection algorithm of Bayesian framework in wavelet domain. Firstly, interesting points are detected and a rough saliency area is obtained by convex hull, then the observation likelihood probability of pixels is calculated by using the color histogram Secondly prior map and the prior probability are obtained by wavelet analysis combining with global and local features. Finally, the posterior probability is counted through the Bayesian theory in order to get saliency value.(3)Applying the proposed algorithms into image segmentation and content perceptual image zooming. Segmentation results are more accurate and the result of image zooming is to keep the shape of the original image. The algorithms use Matlab simulation in the Achanta data sets. By comparisons of ROC curve, precision, recall and run time with the state-of-arts methods, the proposed methods can improve the detection accuracy. The image segmentation is accurately, and the effects of content perception image zooming are satisfied.