Research On Image Annotation Based On Core Semantics

Image annotation is one of the important research directions in the image processing field.With the development of the Internet,the number of images on the network increases rapidly,and it is more and more difficult to find the required images.The image annotation method based on core semantics is more helpful for the retrieval and recognition of subsequent images.This paper mainly focuses on reducing irrelevant semantic labels and locating semantic representation regions of images.Firstly,the image annotation algorithm with irrelevant semantic deletion is constructed in this paper.The category of the object and the number of the object in the image are uncertain,the image classification can not completely label all the objects of the image,and the similarity annotation will produce a lot of irrelevant semantics.Therefore,a method of combining classification annotation and similarity measure annotation is adopted.By obtaining the correlation between semantic tags and deleting the irrelevant semantics.The experiment obtains better experimental results on two different datasets.Secondly,image annotation of semantic representation region is designed.When the human visual nerve touches the image,the brain nerve connected to the visual nerve can find the most expressive position of the image.According to the mechanism of the brain nerve,a semantic representation region labeling algorithm is proposed.The method combines slide window and distance measurement to find the semantic representation region.The similarity measure is used to generate new image annotation for semantic representation region.This method can better mark the object of the core semantic region of the image.Finally,the semantic representation region labeling algorithm based on grid occlusion is completed.In order to obtain a more complete semantic representation region,after dividing the image grid block and then executing occlusion.Through iteration to get the complete semantic representation region,the original image feature and semantic representation region image feature carry out fusion.At last,looking for similarity images for labeling image annotation.Experiments prove that this method can find the complete object feature,and improve the core semantic region marked results.