Research On Mid-level Visual Feature Of Image

With the development of multimedia and computer network technology,it accu-mulates many digital images in the network,which bring new technical challenges.But,unfortunately,most machine learning algorithms are not equipped to handle it directly.Then they make research on finding good visual representations to show the nature of this visual data at the raw pixel level.For this reason,more and more approaches aim to excavate the visual features of this images.Many researchers pay more attention to the mid-level visual feature method for its good performance.It can satisfy two requirements:1)to be representative,it occurs frequently in the tested images;2)to be discriminative,it is different enough from the rest of the visual world.The mid-level visual elements could be parts,objects,visual phrases and so on,but are not restricted to be any one of them.It expresses appearance distribution better than the low-level features,but does not require the semantic ground-ing of the high-level entities.This article mainly studies the mid-level visual features of images.The goal of this paper is to discover the mid-level visual patches in one kind of scenes which can represent special information of the scene well.This paper uses the weak labels to divide sample set into positive and negative,and it never make other labels,such as bounding-box labels.According to the above requirements,in this paper,the main process of the proposed algorithm is divided into two steps.In the first place,it uses Gaussian Mixture Model algorithm to cluster the positive dataset which can obtain the representative information of this set.Then,it splits every gaussian model into several gaussian models by means of bringing in posi-tive sample set and "virtual" negative sample,which can eliminate no discriminative information.Thus,it obtains the gaussian models which have both representative and discriminative.Ultimately,we can extract the mid-level visual elements of the scene,that is,the patches only occur in the positive set but not in the negative set.This approach changes the method of online kernel density estimation with gaus-sian kernels,which has never been appeared in previous research.And that,it is easier to handle and has lower time complexity than before.Ultimately,the experiment has good effect in natural scene graph.