Research On Image Annotation Based On Regional Segmentation

With the development of technology,the number of digital images has increased dramatically.People need efficient management of these images.Thus there is need for automatic image annotation.Most of the previous automatic image annotation methods use low-level features such as color,texture and so on.With the development of deep learning,people use the deep network to annotate the images automatically,and the annotation result is greatly improved.Based on deep learning,this paper attempts to divide the image into several parts in different ways to improve the annotation result.Our main contributions can be summarized as follows.First,we annotate images based on Attention mechanism.In order to solve the problem that the Attention mechanism can't focus on specific area,we propose proposal guiding Attention algorithm.We conduct experiments on the data set VOC2007,the precision increased from 0.66 to 0.69 while the recall increased from 0.49 to 0.60 and the F1 indicator increased from 0.56 to 0.64.The result has been significantly improved.Second,we generate proposals using unsupervised method and mark the images in a targeted way.Then we fuse the whole map label and get the label of the whole image.We conduct experiments on three public datasets,VOC2007,VOC2012 and MIRFLICKR-25 K and the MAP values are 92.29,91.09 and 74.14 respectively,which leads the field.Compared with the similar algorithm HCP-2000 C,on dataset VOC2007 and VOC2012,our MAP values increased by 7.09% and 6.89% respectively,which fully proves the effectiveness of our method.Third,in remote sensing image annotation,different from the conventional low-level features,we use deep network to extract the features of remote sensing images and propose a new method of image dividing.Taking into account objects in remote sensing images have different size.We cut some pictures from high-resolution remote sensing images of Vietnam to conduct experiment and achieve good results.Compared with the previous method,we can better annotate smaller targets.We try different segmentation methods on natural images and remote sensing images,so the annotation can be targeted.The research results have good theoretical and practical value.