Research On Image Ranking Model Based On Preference Relations

With the rapid development of Internet technology, electronic products are morehigh-end, and since the mobile phone, digital cameras and tablet PCs etc. are available,pictures can be taken in various ways. At the same time the upload technology and sharingtechnology make more images, videos etc. continue to spread on the Internet. It hasbecome an issue of growing concern that how to organize and manage with these imageseffectively. In view of this, on the basis of analyzing the related domestic and internationalresearch results, this paper studies the image ranking model, and some image sortingalgorithm are proposed.Firstly, the model of relative attributes and a sort algorithm of crowd density based onrelative attributes is presented. As opposed to predicting the presence of an attribute, arelative attribute indicates the strength of an attribute in an image with respect to otherimages. First of all, images in training set are scored and ranked, and then a rankingfunction of crowd density is learned by a set of images pairs with preference relationscombining with features extracted from the training samples. The last step is to score andrank for images in the test set. The experimental results show that this algorithm of thesort algorithm has excellent performance.Secondly, a personalized relative beauty ranking algorithm is presented based on asubject’s personal taste. Firstly, training data of faces are sorted based on a subject’spersonal taste. To avoid the high cost of sorting, a sorting method that includes binning isused. The second step is to extract SIFT features and HOG features from images oftraining set, after that three personalized relative beauty ranking functions are learned bySIFT features, HOG features and a blend of the two features combining with a fullyordered list. Lastly, images of the test set are scored and ranked, and the value of KendallTau of the ideal results with sort results is to determine the accurace of attractivenessranking method.Finally, a novel multiple-instance images rank learning framework, Average andSoftmax multiple-instance ranking algorithms are proposed based on the multiple-instance learning. First of all, the images in training set are divided into blocks. Each image isregarded as a bag and blocks are regarded as instances of bags in multiple-instancelearning, then instances are scored. Average algorithm treats the arithmetic average valueof scores of all instances as the score of bag. Given different weights with differentinstances (the higher score the larger weights), Softmax algorithm treat the weight sum ofscores of all instances as the score of bag.