Research On The Data-driven Novel View Image Generation Methods For Armored Targets

Tank armored targets are the main forces in ground operations and have important military value.In wartime data preparation,it is necessary to analyze the images of adversarial targets in various views for priority judgment or strike position selection.The antagonism of war determines that obtaining complete intelligence information of military targets,especially adversarial targets,is a difficult or even impossible work.In order to make up for the missing view data and expand the complete data samples,this paper studies the novel view image generation method of armored targets.It has the following difficulties: 1)the corresponding dataset is missing,and the evaluation index is single;2)under the background of special application,the view transformation span is large,and the number of input source images is small;3)there are some problems in the actual scene,such as the lack of view label,background interference and so on.Therefore,in order to improve the applicability of the model to complex scenes and problems,this paper focuses on data-driven armored targets novel view image generation method research.Firstly,aiming at the problem of lack of corresponding dataset and single evaluation index,this paper constructs a set of datasets suitable for novel view image generation of armored targets,and puts forward new evaluation indexes for classification and recognition.Combined with the research background and research objectives,this paper prepares the 3d Max simulation dataset and real shooting model dataset of armored targets.Considering whether the data generation results can contribute to the improvement of target recognition network performance is the key to test the advantages and disadvantages of the generation technology of this subject,this paper not only measures the difference between the generated image and the truth image at the data level,but also puts forward a new evaluation index at the application level.Secondly,aiming at the problems of large view transformation span and small number of input source images of non-cooperative targets,a novel view image generation network based on convolution block attention module and confidence map is proposed.The convolution block attention module is used to improve the feature extraction ability of the network,and the prediction results of pixel generation network and flow guidance network are combined with confidence map to deal with the input of any number of source images.In this paper,the performance of the proposed algorithm is verified on the armored target 3d Max simulation dataset.The experimental results show that the novel view image generation network based on convolution block attention module thinning feature and confidence map weighting fusion can generate clear texture details and ensure the rationality of the whole structure of the target object.Finally,aiming at the problems of missing view labels and background interference in the actual scene,a novel view image generation model based on spatial transformer network and target segmentation is proposed.The spatial transformer network is used to reduce the sensitivity of the model to the angle,the novel view image generation of armored targets is realized without using the accurate view label.The region of the target is located through the saliency object detection network,and on this basis,a target segmentation module is proposed to eliminate the interference caused by background noise.In this paper,the performance of the proposed algorithm is verified on the armored target 3d Max simulation dataset and real shooting model dataset.The experimental results show that the novel view image generation model based on spatial transformer network and target segmentation can basically solve the problems of missing view label of input image and background interference.