Research On Video Object Segmentation Algorithm Based On Learning Attention Modulation Network

With the rapid development of information technology and the rapid popularization of the Internet,video,as an important information transmission carrier in the Internet,has deeply affected the development of life,education,social contact,military and other fields.As one of the key technologies in video processing,video object segmentation has been widely used in intelligent video monitoring,intelligent traffic,video editing,unmanned driving and other fields.In the actual scene,the existing video object segmentation algorithm is easily challenged by such factors as object occlusion,illumination change,rapid movement,deformation,etc.,which makes the model unable to accurately locate the segmentation target,resulting in problems such as missing target or incomplete target contour,resulting in low segmentation accuracy and poor robustness.This paper focuses on the research of semi-supervised video object segmentation algorithm.The goal is to design a robust segmentation model that combines temporal information and spatial relationships for network modulation,and can quickly adapt to specific segmentation object instances.The main research work of this paper is summarized as follows:This paper proposes a semi-supervised video object segmentation algorithm based on feature attention modulation network.The visual attention network and spatial attention network are constructed to learn the semantic information and spatial information of segmentation objects respectively and modulate the segmentation network to focus on specific object instances.In this way,it can not only accurately represent the appearance information of the segmentation target,but also make use of the motion information of previous frames,so that the segmentation model can learn the characteristics of the segmentation target more robustly in complex scenes.In order to make better use of multi-scale features,the feature attention pyramid module is proposed to capture multi-scale feature information by pooling nuclei of different scales.The experimental results show that the proposed method is robust and accurate,and achieves high quality segmentation under complex conditions such as object deformation,rapid movement and scale change.Aiming at the challenges of occlusion and deformation between similar objects in the multi-object segmentation scene,this paper proposes a semi-supervised video object segmentation algorithm based on double attention modulation network,which is further expanded on the basis of work 1.The channel-space attention module is constructed to further refine the modulation segmentation network and enhance the features related to the segmentation object,so that the segmentation network not only focuses on the segmentation of the overall information of the object,but also focuses on the local features.This paper also proposes a residual refinement upsampling module,which combines semantic information of high level with location information of shallow level to realize the fusion of multi-scale features.In order to solve the problem of unbalanced training samples,focal loss function is also used in this paper to make the network focus on the training of difficult samples and accelerate the convergence of the network.Aiming at the problem of model overfitting caused by insufficient training samples,"lucid dreaming" data enhancement strategy was adopted to simulate the changes of future frames and generate a large number of training samples.Experimental results show that the model proposed in this paper achieves excellent results in multi-object segmentation tasks?...