Key Technology Research Of Augmented Reality Based On Deep Learning

Augmented Reality(AR)technology is a research hotspot in the current computer application field.Its essence is to superimpose virtual information or objects in real scene images,so that virtual objects can be integrated with real scenes to enhance people's perception of the real environment.Interact with the experience.It combines digital image processing,computer vision,intelligent pattern recognition and many other technologies.With the development of various related technologies,augmented reality has become more and more widely used.However,there are currently three technical difficulties: target detection,scene perception and 3D registration,image display and human-computer interaction in complex scenes.Target detection is the first step to realize augmented reality system.In the real environment,there are many types of targets,which are easy to be occluded,the volume is very different,and the background and texture are complex.The current processing method is difficult to accurately extract the key information of the target,resulting in low detection accuracy,and can't meet the requirements of augmented reality display and update.In the scene perception stage,the traditional depth map acquisition method is insufficient in adaptability and robustness.It is easy to cause matching problems due to the diversity of scenes,which directly leads to the inaccurate registration of virtual and reality,which directly affects the user experience.In the case of fast motion and large-scale moving scenes,the establishment and display of 3D models are not ideal,which directly affects the human-computer interaction experience.In view of the above problems,this thesis studies the complex scene target detection,scene perception and 3D face modeling methods in augmented reality applications.The main work is summarized as follows:1)Aiming at the problem of complex scene target detection in augmented reality applications,a complex scene target detection method based on deep learning is proposed.This method designs a bidirectional pyramid fusion network to better extract complex characteristics of the target in the background environment through a series of bidirectional convolution operations.At the same time,the relative area recommendation network is used to extract the target features and local features of different scales to the maximum extent.In addition,the method incorporates information containing the target context to further improve the detection performance of the target in complex scenes,especially for scenes with different occlusion and scale.Experimental results show that the method has high detection accuracy.2)For the adaptive and robustness of scene perception in complex scenes in augmented reality applications,it is easy to cause matching due to the diversity of scenes.This paper proposes a depth image acquisition method based on spatiotemporal consistency constraints,and based on the scene,a sequence of binocular video images is to perceive a three-dimensional scene.Based on the depth information of each frame,the method combines depth value information,segment information and optical flow field information in the time domain to optimize the initial depth of each frame and enhance the depth of the video sequence depth map in the time domain,the continuity and consistency of the value changes,thereby reducing distortions such as noise,tailing,and flicker,and improving the augmented reality display.3)Aiming at the 3D model establishment and display,combined with the project background,taking the most common face model in the application as an example,a low cost,high precision and high robust face modeling method is proposed.This method generalizes the traditional monocular bilinear model method to the multi-view bilinear model,combines the a priori feature constraint of the multi-view image with the texture constraint,and uses the prior feature constraint as a significant prior condition to estimate the accuracy.3D facial contours and texture constraints are used to obtain high-precision 3D face shapes,which improves modeling accuracy compared to traditional methods.At the same time,the method fully exploits the three-dimensional information implied in the multi-view image and improves the robustness of the result.