Research Of Image Saliency Based On Reciprocal And Spectral Residual And SLIC Super Pixel Segmentation

With the rapid development of Internet technology, tens of millions of image information has been received. It is impossible to meet the needs by the human being to deal with the image information, people prefer to use the computer to process the image information. The researchers simulated the human visual attention mechanism and proposed the image saliency detection technology. However, there are still some defects in the image saliency detection methods, such as accuracy, integrity and computational complexity. In this paper, we introduce some typical regional detection models, point out the shortcomings of the existing methods and propose two improved methods. The main contents are as follows:(1) In order to improve the shortcomings of the existing detection methods,a salient object detection method based on reciprocal and spectral residual is proposed.First, the proposed method uses the difference between the gray image and its corresponding Gaussian low-pass one to extract the image of the colour feature ,meanwhile it further reduces the number of Gaussian pyramids by 6 levels in order to avoid redundant;second,a reciprocal function filter is used to extract local orientation information; third, spectral residual algorithm is used to extract spectral feature;finally ,three extracted features are properly combined to generate the final saliency map. The experimental results based on the two mostly common benchmark datasets show that the proposed reciprocal function and spectral residual method are significantly improved in the three indicators, precision, recall and F-measure.(2) In order to improve the segmentation of RC method, a significant area detection method of SLIC super pixel segmentation is proposed. First, three different scales of images are obtained and the SLIC super-pixel segmentation is carried out for images, Second, the spatial weighted contrast and position information feature of each scale are calculated;finally, we allocate significant values for the region and combine the significant values for each layer.The generated maps not only have clear boundary, but also can effectively restrain the influence of the complex background,and improve the performance of the detection method.