Research And Implementation Of Image Matching Algorithm Based On Deep Local Feature

As a basic problem of computer vision,image matching is widely used in pose estimation,3D reconstruction and image retrieval.However,in the process of image matching,it is easy to be affected by various external factors.The traditional image matching algorithm based on handcrafted local features have fast inference speed,but poor accuracy in the face of bad lighting and viewpoint conditions.The image matching algorithm based on deep learning is limited by the number and discrimination of training datasets,so it has poor effectiveness in practical application.In view of the limitations of the existing two methods,our paper suggested an image matching algorithm based on deep local feature,and proposed an image recognition method for Web AR scene,which verified the practicability of the algorithm.After all,the main work of our paper is as followed:(1)Aiming at the limitations of existing image matching algorithms,our paper proposes an image matching algorithm based on deep local feature.The image matching pipeline mainly concludes three steps:keypoints detection,descriptor generation and outlier filtering,and we completed these steps in different ways.The traditional handcraft feature and deep local feature are combined to improve the robustness of image matching algorithm to illumination,perspective and rotation.In the IMC-PT dataset released by CVPR 2020,we use multiple evaluation metrics such as mAA to complete the test.The results show that the proposed algorithm has better accuracy and robustness than SIFT,D2Net and other algorithms in different lighting and viewpoint changes.(2)Aiming at the problem of background independent matching point interference in image capture,our paper proposes a saliency detection model based on lightweight network.Saliency detection algorithm can locate and segment the main part of the image,extract the region of interest from the user's high-resolution image,and remove the irrelevant clutter background.The experimental results show that the volume of the lightweight saliency detection model proposed in this paper is 94%smaller than that of the latest RAS model,so that the model can be deployed and run in the device browser,used for device-side preprocessing,speed up the back-end server image feature extraction and matching,and improve the utilization of server computing resources in the project.(3)Based on the above two algorithms,this paper designs a Web AR image recognition system based on cloud-edge collaboration.Image recognition is a crucial step in the whole Web AR process.The system designs the device side preprocessing module based on the lightweight saliency detection model,designs the image recognition module in edge and cloud server based on the image matching algorithm,and realizes the cloud edge cooperation through the KubeEdge framework.The practical application verifies the value of the algorithms proposed above.It provides a new idea for image recognition in Web AR scene.