Multispectral Image Matching Algorithm And Implementation Based On Domain Adaptive Siamese Network

As one of the key technologies in the field of computer vision,image matching technology has been widely applied in aircraft navigation and remote sensing reconnaissance.It is a classical difficult task to realize reliable matching between images acquired from different time and viewpoints in complicated scenarios.Especially in the aspect of matching multi-spectral images obtained by unmanned flight platforms,it is difficult for the traditional matching algorithm to achieve universality due to factors such as distinct imaging characteristics of different spectrum and wide search area.In this case,this paper mainly focuses its research on how to effectively extract the common features of multi-spectral images and accurately locate similar areas,improve the algorithm's antiinterference ability and accuracy,and finally their applications on real-life scenarios.Considering matching multi-sensor images,there is a precedent of heterogeneous siamese matching network,which requires a massive number of data training on heterogeneous spectral images under the same scene,so that it is difficult to be really applied in real life scenarios.In order to construct a model on the common features of multi-spectral images based on various spectral image data in different scenarios,this paper designs a domain-adaptive feature network DASiam Net in multi-spectral images.Based on the consistency of imaging in the same spectrum and the difference between the imaging of multiple spectrum,the channel attention mechanism and the domain attention mechanism are introduced.Before the heterogeneous siamese network structure,a domain allocation channel adjustment unit is designed to adapt to the data input of different spectrum,and combined with the CIR units to realize adaptive feature extraction of multi-spectral images.Based on multi-spectral feature extraction feature network,this paper designs a multifeature fusion module in the pyramid structure,and proposes a domain-adaptive siamese matching network DASiam RMN with RPN structure,thus achieving adaptive matching of multi-spectral images.To further compress network model parameters and reduce algorithm complexity,and make the matching network truly an end-to-end network,this paper draws on the idea of center point detection in Center Net,and constructs a central matching domainadaptive siamese matching network DASiam CMN.CMN matching detection network is adopted to turn point-by-point traversal matching search into a central point detection problem to achieve reliable matching in a wide range of scenarios.The multiple networks and series of siamese networks composed in this paper are comparatively tested on public data sets and custom hybrid multi-spectral data sets.Experiments show that DASiam Net,acquired through open data training,is adaptive to multi-spectral data input,performs well in cross-spectral matching tests,has good generalization capability and less dependence of heterospectral images on the training samples.Domain-adaptive siamese matching network DAsiam CMN can effectively adapt to large scale searches and matching scenes with dramatic changes in angles,and does not depend on the setting of artificial anchor frames.What's more,there is no subsequent calculation process of non-maximum suppression,so it's quite suitable for embedded devices.Finally,this paper selects the Hi Ascend310 processor to cut and transplant the matching network models,verifying the algorithm's feasibility on application.