Research On Feature Point Detection And Descriptor Extraction Based On Deep Learning

Feature point detection and descriptor generation are the basic contents of image research.Image feature point detection is an essential part of many tasks,such as image retrieval,object recognition and so on.Excellent image feature points can guide the extraction of image features with rich detail information,which is conducive to the performance improvement of subsequent tasks.After feature points are detected,local image information is represented by image descriptors.Image descriptors can express the features of image in the form of digital quantization,and it is convenient for storage,transmission and processing.Excellent image descriptors are robust to features and have great significance for tasks such as automatic driving and 3D reconstruction.Therefore,this paper proposes a new feature point detection network and a new descriptor generation network.The main work of this paper is as follows:Firstly,the classical algorithms of feature point detection and descriptor generation are introduced.The hand-crafted Harris algorithm,SUSAN algorithm,DOG algorithm,HOG algorithm,SIFT algorithm,SURF algorithm based on manual design and the learning-based algorithm Quad-Network and Hard Net are analyzed in detail.This paper outlines the current status of feature point detection task and descriptor generation task.Secondly,a feature point detection network based on differential response graph is proposed.The network is based on an unsupervised training method,which uses random non-affine transformations to generate true annotations of the training data to train the network to be invariant to the corresponding transformations.At the same time,the network uses random affine transformation to augment the data to avoid training overfitting.In addition,the proposed network uses differential convolution kernels to replace traditional convolution kernels to extract image features,generating feature maps with more detailed information.Compared with the classic algorithm on the relevant data set,the experimental results show that the proposed network has shorter training time and improved detection performance.Finally,a new descriptor generation network is proposed.In this network,Pearson correlation coefficient is used to evaluate the correlation between descriptors,and a new network training method is proposed to generate better-performing image descriptors.The proposed network is superior to the classic SIFT algorithm and the currently advanced L2 Net and HardNet in test results,and has a positive significance for performance improvement.