A Study In Semantic Image Annotation Based On Local Model

With the rapid development of the internet, we are faced with the rapid increment ofall kinds of multi media data including image. How to retrive these images effectively isan urgent need. However,low-level features can’t express the semantion of the image, wewant to retrive images based on keywords. Automatic image annotation is one of the keyto solve this problem. This paper focus on image annotation methods based on local modelwhile taking label occurrance into consideration.Three different kinds of local models are implied including a knn based method, alocally regression model, and a locally classification model. Severel method areconsiderated using different setting in those models. In knn based methods, non weight,distance based weighted and rank based weighted methods are expiremented respectively.In regression models, semantic space is defined and image annotation is viewed as aprocess mapping features into the semantic space. Statostic gradient descend is used foroptimion. A semantic locally regression model is achived by taking label occurranceinformation into consideration. In classification models, multi-label problem are turnedinto two classes classification and solved using SVM. Semantic kernel and a explicitlypost process considerate the label occurrance are studied.Mass of expirements are conducted and the results of which suggest that buildingmodels locally can improve the performance. Some methods in this paper almost achivedstate of art performance. The unbalance of positive and negative instance during trainingOVR classifier is a severe problem and reweight should be considerate. Taking semanticinformation of label occurrence into consideration doesn’t improve the performancenecessarily, while it increase the model complexity centarnly.