Natural Images Segmentation Based On Active Contours Models And Saliency Map

Images are an important method to get and exchange information for human beings. In the computer field, the main purpose of digital image processing is to get objective information and establish the relevant description of the image by analyzing objects in the image.Image segmentation is a fundamental computer vision technology and a basic premise for image analysis and pattern recognition. Due to natural images are vulnerable to complex background and irregular change of illumination, it is difficult to obtain stable segmentation results. In the image segmentation area, to get an accurate description of image's objects have been studied for long time. Generally, an image contains information of texture, color, profile and others, which means that in the segmentation process we should consider the local and global factors. Based on review and conclude the relevant domestic and international literatures, we have done a serious research and analysis about image segmentation approaches. Then, we have explained exist problems, background and significance of natural image segmentation. According to the characteristics of natural images and its own special needs, a detailed study of image segmentation method which is based on active contour model and a mechanism of visual attention analysis have been done.Level set model is sensitive to noise and prone to false edges, which lead to stop evolution because of curve evolution into a local optimum and issues of uneven region and sensitivity of initial contour. Affected by active contour model based on region, we proposed an improved geometric active contour model which introduces the whole region as a driving force into energy function. This model uses statistical information-driven curve evolution which will effectively protect the boundary information and make the curve evolution stop at the boundary position which leads to improve the segmentation accuracy at certain extent. Since the whole evolution process does not require re-initialization and does not need to calculate the signed distance function, the calculation is greatly reduced. The segmentation results will be different when we use different initial segmentation position in that the model is vulnerable to the initial level set contour position.If the initial contour can advance around the near object, it can drastically reduce the number of contour evolution. Therefore, based on the improved geometric active contour model, we proposed an approach which is based on the analysis of whole frequency and intensity normalized saliency map by introducing visual saliency map analysis mechanism for the initial contour extraction. The output of our method preserves the full resolution attention of the object boundary which is better than the traditional methods and can better extract the image edge information. Then, we get the initial contour by using the attention image which segment by OSTU-based adaptive threshold method. Experiments show that this method increases the speed of image segmentation while also avoiding over-segmentation of the image and can better extract the target area. Compared to other active natural image segmentation methods, our method has made great progress in the segmentation results, speed and complexity.