Research On Technology Of Extracting Image Salient Region

Vision is the most important way which people perceiving and understanding the world, it’s consisted of the stimulation which came from the nervous system when a large amount of information was obtained through the eyes. Visual saliency is the computer simulation of human vision, it could rapidly and precisely locate a target or interesting region from the complex scenes. In short, it’s the description of the ability that objects attract visual attention.There are a lot of redundant information in digital images which were known as the main visual carrier in computers. The most important content in a image is often only focused on some small regions. These regions were called salient region because it attracted more visual attention and represent more contents than others. Detection, extraction and representation of these regions has a widely application value. In recent years, with the development of the computing models on visual saliency, it has become a research focus.This paper mainly introduces the technology of image salient region extraction. The research includes the following content:1.A new method based on global comparison for extracting of image salient region. We use an adaptive size image segmentation technique cut image into blocks, take colors, block similarity and spatial positions as the basic feature, calculate the saliency for each pixel under the multi-resolution images. After then, a interpolation algorithm was used for image fusion and normalization. At last, the salient regions were obtained by threshold from the normalized saliency map. Since the procession of calculating the saliency map in our algorithm have split and merged blocks, then we don’t need any segmentation more. When compared to the two kinds of existing algorithm, our algorithm is simple and fast. In the mean time, it has a low computational complexity. The experimental results show that our algorithm were in accordance with human visual attention mechanism and had a higher match rate with reference template.2. A new method based on multi-scale filtering in frequency domain for extracting of image salient region. Firstly, we got the multi-scale map for input image through the different filters in frequency. Different resolution choose the different kinds of filters. In the high dimension space, we use Gaussian Laplacian filter to keep clear contour information. While a spectra different mean filter were used in the low dimension spaces. And then a weight function was introduced to normalize the saliency map of each resolution. Finally, the salient regions were obtained from the normalized saliency map by a modified K-means segmentation algorithm. Comparing with the results of the existing frequency algorithm, our algorithm have a better significant area and a more clear contour. The precision and recall rate from the extraction results were also mucher higher than others.