Research On Image Deblurring Algorithm Of UAV Based On 3d Reconstruction Application Scenario

Large-scale urban 3D reconstruction of scenes has a wide range of application prospects in urban construction and other fields,but the image acquisition process will inevitably be affected by environmental factors,camera shake and other factors,resulting in image degradation,and then affect the reconstructed architectural model.Therefore,the restoration of fuzzy image is crucial to the effect of 3D reconstruction.Fuzzy image enhancement is also an important research topic in the field of computer vision.The traditional deblurring method usually needs to estimate the fuzzy kernel and design the feature model manually through a large number of image priors,but it is often not generalized,and the restored image still has ringing effect and noise.At present,the main problem of deblurring algorithm based on deep learning is that it relies heavily on data set,but the method of data collection is limited,and the detailed texture information of restored image is not rich enough and there are many problems such as imperfect details of the restored image and blurred image edges.Aiming at the problem that it is difficult to collect data sets that meet the conditions,this paper proposes a three-level fuzzy image data set expansion method based on depth learning in combination with the background of urban 3D reconstruction,thus realizing the expansion of fuzzy image data set.In addition,aiming at the problem of detail loss of restored image,a generation antagonism network structure based on residual attention mechanism is constructed,which improves the restoration of many details in the image and the texture of the building frame.The research work on these two key issues is as follows:(1)Proposes three methods of fuzzy image data set expansion based on depth learning.First,analyze the possible fuzzy interference in the process of UAV image acquisition,and finally determine three kinds of blurring that will inevitably occur,namely,motion blurring,atmospheric turbulence blurring and defocusing blurring.Then,construct three kinds of blurring respectively for the real stadium building image set,and mix the three different kinds of fuzzy images with the original clear images to form the training data set.The generated model is tested with the model that only constructs a single fuzzy dataset to train and generate.Through comparative experiments,the data set expansion method proposed in this paper can train a better model and achieve better image deblurring effect.(2)A method of generating antagonism network based on residual attention is proposed.In view of the phenomenon that most of the deep learning deblurring algorithms have lost details and the edge contour recovery is not obvious,this paper improves the generation of the confrontation network by adding the residual attention module and the weighted residual dense block on the original basis.The backbone network uses the feature pyramid structure to extract the image features,including different layers from coarse to fine,At the same time,the residual attention mechanism module is added to process the extracted information,which can adaptively learn the image edge and image spatial structure.In addition,adding weighted residuals dense blocks can compensate for the information loss caused by multiple sampling operations in the backbone network.The experimental results show that this method has achieved better experimental results in both subjective and objective measures for the restoration of aerial degraded images.