Research And Application Of Deep Interactive Image Object Segmentation Algorith

The interactive image segmentation method can obtain segmentation results that meet the user’s intention under limited user interaction.It has been widely used in image editing,data annotation,automatic driving and other fields,and has important academic value and broad application prospects.With the rapid development of deep learning technology in the field of computer vision and image processing,how to make deep networks to better understand the user’s segmentation intention and improve the user’s interactive segmentation efficiency has also become a current research hotspot.This article focuses on the research of interactive image segmentation methods under the framework of deep learning,explores the tasks of user-interested object segmentation in the convenient mouse-click interactive mode,and addresses the issues of ignoring the difference among interactive behavior,lack of correlation between interaction sequences and more.The specific research work is as follows:(1)The mainstream iterative click training strategy treats each interaction equally in the same way,ignoring the unique intentions and effects of each click,and dragging down the user’s interactive segmentation efficiency.In response to this problem,the entire interaction process is divided based on the user’s interaction habits,and the iterative global selection and local correction network model with clear goals is proposed.Through the feature map mapping,loss function definition and multi-scale interactive intention attention module design,the network’s ability to recognize different user intentions is enhanced to better mining the intent of each click interaction.The comparison of experiments on multiple mainstream public data sets verifies the superiority of the selection-correction network in the understanding of user interaction intention,the preservation of target edge details,and the efficiency of interactive segmentation.(2)The lack of correlation between consecutive interactive sequences leads to the "forgotten phenomenon" of segmentation in existing deep interactive segmentation methods,that is,the correctness of the previously segmented area cannot be guaranteed in the subsequent interactive process,which seriously affects the interaction segmentation algorithm performance.In response to this,each interactive segmentation process is regarded as a spatial problem,and the continuous interactive correction process as a time domain problem,we design a temporal network to serialize the segmentation clues of spatial changes in sequence to ensure that the entire interactive process can maintain time and space consistency.In order to facilitate quantitative experimental analysis,we have defined a quantitative index used to evaluate the degree of occurrence of "forgetting phenomenon"-the forgetting index.Comparative experiments in multiple mainstream public data sets verify that the above-mentioned time-series modeling strategy can effectively suppress the occurrence of "forgetting phenomenon".(3)With the support of the above content,an interactive segmentation system based on spatio-temporal consistency selection-correction network is designed,including five modules:data reading,hyperparameter configuration,human-computer interaction segmentation,index evaluation,and result export.This system applies the algorithm of this paper to the actual field.