Automatic multi-object extraction from a natural image based on saliency map
Huawei Tian, Yanhui Xiao, Wengang Feng, Jianwei Ding, Yunqi Tang · 2017
In recent years, saliency maps previously computed from images have been employed for unsupervised object segmentation. The performance of existing method is good at single-object extraction. However, these method is powerless for multi-object extraction. An automatic multi-object extracting method is proposed in this paper. Firstly, both of the color feature and texture feature are used to detect the saliency map of image. Secondly, the saliency map is divided into 3 parts to constitute the trimap according to the saliency value. As the coarse trimap, the saliency map is inputted to K nearest neighbors (KNN) alpha matting. The KNN matting will produce the multi-object extracting results. Experimental evaluation on benchmark dataset indicates that the proposed multi-object extracting method is comparable or of higher quality than the existing method.