Research On Multi-perspective Light Field Reconstruction Method For Maneuvering Target Recognition

In recent years,with the rapid development of deep learning,the research direction of image processing has attracted much attention.Object recognition technology refers to the process of using computers to process and analyze images to distinguish an object from other objects.Due to the interference of occlusion,distortion,and blurring,the accuracy and robustness of target recognition is poor.Compared with traditional imaging,the light field contains all the images of the same object taken at different positions and different angles,that is,the light field contains more extensive image information and higher image quality.It can be seen that the light field can be used as the target's natural feature library.Due to the sparse nature of the light field,the reconstruction of the light field greatly improves the application of the light field in practice.However,the limited visual range of the light field and the huge amount of data during the reconstruction of the light field and the complicated calculation process greatly affect the accuracy and real-time performance of the light field in target recognition.Agriculture is the first industry in China,and its development is closely related to people's food and clothing.The actual agricultural scene is usually very complex,different weather,region,soil,light and crop varieties have brought great challenges to the relevant research.Aiming at the problems of limited light field visible range,large amount of light field reconstruction data,and difficult to obtain available data,this study uses multi-intelligence cooperation mechanism to effectively estimate the state of the maneuvering target and solves the problem of maneuvering target loss.In an environment with occlusion and low resolution,the basic data required for image reconstruction cannot be completely obtained.Aiming at the problem that the total number of captured data samples is large in the multi-perspective light field reconstruction process,but there is less available data,the target data is generated or enhanced by using a generative adversarial network to solve the problem of missing samples.Aiming at the complicated calculation of multi-perspective light field,the transfer reinforcement learning model is used to reduce the reconstruction time based on increasing the number of reconstruction samples.The specific content of the method proposed in this article is as follows:(1)A maneuvering target recognition method based on multi-perspective light field reconstruction is proposed.This method firstly represents the light field as a sub-light field(i.e.,a multi-perspective light field)at different views,and performs sparse representation and reconstruction;second,From the perspective of multi-agent cooperation,combined with the distributed collaborative theory of graph theory,a collaboration mechanism between multi-perspective light fields under the constraint of full area coverage is established.(2)A multi-perspective light field reconstruction method based on GAN is proposed.This method first introduces multi-agent cooperation mechanism into the field of light field reconstruction,and effectively removes redundant and High-noise reconstruction of sample data ensures the continuity of maneuvering target monitoring.Secondly,according to the characteristics of GANs that can generate and enhance data,the sample data required for light field reconstruction is improved,and a GAN-based multi-generator and the discriminator model is used to complete the reconstruction of the light field.(3)A multi-perspective light field reconstruction method based on transfer reinforcement learning is proposed.This method first establishes a similarity measurement model and selects reinforcement learning or feature transfer learning models based on the similarity threshold.Second,it establishes reinforcement learning models that use multi-agent Q learning.Determine the feature set that is similar to the target domain and feed it back to the source domain to increase the labeled sample capacity.Finally,establish a feature transfer learning model and use PCA to obtain the maximum feature embedding space of the source and target domains and use it for label data.migrate.(4)In different scenes,the proposed light field reconstruction algorithm was verified by target recognition simulation and compared with the existing methods.In addition,the validity of the method proposed in this paper was verified by the significance hypothesis test.