Research On Image Feature Point Extraction Algorithm Based On Neural Network

For decades,people have made in-depth research on feature point extraction algorithms,and proposed a number of successful feature point extraction algorithms,which provide a basis for image matching,image registration and target tracking in the field of computer vision.The best techniques available today are still based on hand-crafted features.Although hand-crafted feature point detectors work well,they often rely on pre-designed structures and are not very flexible in different situations.With the development of deep learning and neural networks,applications based on neural network technology have begun to be competitive in various fields.Aiming at the limitations of traditional feature point extraction algorithm,this paper proposes an image feature point extraction method based on neural network to improve feature point detection ability.The main research content of this paper is as follows:Firstly,this thesis deeply explores the traditional feature point extraction algorithm and the structure of neural network,analyzes the limitations of traditional algorithms and the excellent generalization learning and nonlinear mapping ability of neural networks,and proposes an image feature extraction method based on neural network.Secondly,this paper conducts an in-depth study on several classical convolutional neural networks and summarizes the development of convolutional network models.In-depth exploration of an existing self-supervised learning feature point detection model,and proposed a method to improve its shared convolution layer.Based on the structure of the VGG(Visual Geometry Group)network model,four improved network models are given.And verify the ability of the network model to extract feature points from the image under varying angles of view and light changes on the HPatches dataset.The experiment proves that the 11-layer network structure has achieved good results,and the repetition rate of the feature points extracted under the change of the viewing angle and the light change is increased by 0.9% and 1.3%,respectively.Finally,the feature point descriptor is trained by the 11-layer network model,and the matching ability of the feature points under the change of the angle of view and the change of the light is tested.At the same time,compared with traditional feature point extraction algorithms such as SIFT algorithm and ORB algorithm.Experiments show the effectiveness of the improved method.