Research On Image Multi-labeling Algorithm Based On Spatial Information And Migration Learning

With the rapid development of Internet technology,the demand of network information has been growing exponentially.In addition to the requirements of text information,people pay more attention to the cognition of the image content.Automatic image annotation can compensate the shortages of wasting much time and effort comparing with manual annotation.This method improves the efficiency of image understanding technology.But now people's understanding of the image content has not only rigidly adhere to a single concept and mark,it is inclined to multi-level and multi-angle interpretation.So image multi-label learning has emerged,which better adapts to the needs of people.The multi-label learning methods of images emerge in endlessly and gradually mature,the application of image spatial information has becoming more and more fully.But in the real world,in addition to the complete images,there are a large number of incomplete or obscured images,which also contain a large number of effective information.Based on this part of the special image groups,this paper proposed an incomplete image annotation algorithm based on spatial information by using spatial information.The experiment result proved that our method could reduce the negative impact of image content understanding which is influence by the picture incomplete part.This kind of method could also improve the recall ratio and precision ratio of image annotation.It could reflects-hidden information of the whole image better.At the same time,the correlations between different images,images and marks,different marks will need to be more fully utilized.In this paper,we proposed an image annotation scheme based on similarity transfer learning to explore the correlation between images and tags.Experiments result show the effectiveness of our image annotation method.This kind of method could improve the quality of image annotation,and reduce the influence of the noise of images.The main work and innovation of this paper are summarized as follows:1.Aiming at the incomplete image groups,,we proposed an incomplete image annotation algorithm based on spatial information,which combined with the importance of the spatial information for understanding the image content,.We first selected the minimum rectangular region of the incomplete part of the picture.And then we segmented all the images by extending the selected region as the segmental directrix.We calculated the similarity of pictures by using the image segmentation sub-block as the fundamental unit.We used the spatial structure information of image segmentation sub-block to complete the automatic image annotation.This method could reduce the negative impact of image content understanding which is influenced by the incomplete part of the picture.Our method could also improve the recall ratio and precision ratio of image annotation.It could better reflect the information contained in the whole image.2.To further explore the correlation between images and tags,we fused migration learning theory into image multi-label learning method.We proposed an image annotation scheme based on similarity transfer learning.First of all,we established similarity measures between features of images.Then we introduced the thought of similarity transfer learning which could transfer the similarity of the image characteristics to the similarity of the image annotation.After that our work realized the image annotation by using the statistical methods.This kind of method could improve the quality of image annotation,and reduce the influence of the noise of images.This correlation can provide additional useful information for learning to a certain extent,and it also can compensate the deficiency of the sample data.We used the regional structure information and fused the migration learning theory into image multi-label learning method.Multi-label learning algorithm proposed in this paper can improve the mark performance of incomplete image population and has great robustness.For complete image,the algorithm proposed in this paper can effectively weaken the interference,and it also can enhance the mark learning effect.