Research On Feature Point Extraction And Description Algorithm Based On Deep Learning

With the development of artificial intelligence and autonomous driving,important tasks in the field of computer vision,such as simultaneous localization and mapping(SLAM),structure from motion(SFM),camera calibration and image matching,have received increasing attention.The accuracy of these tasks dependes heavily on the performance of local image features.In previous work,a large number of handcrafted features,such as SIFT and HOG,etc.,were used to extract image features and describe them.However,in the real environment,the existence of changes in image perspective,scale,and illumination has brought great challenges to traditional handcrafted features,which needs to be improved in terms of performance and robustness.On the other hand,the continuous upgrade of computer hardware and the rapid development of computing capabilities have driven the revolution of deep learning,which has made breakthrough progress in visual tasks such as object detection,semantic segmentation,and depth estimation.In these works,a large number of manual annotations are usually used as the true value.But in the feature points annotation,this is still a ill-conditioned problem.At present,most feature point extraction methods based on deep learning still use traditional handcrafted features as true values.The distribution of feature points and descriptor of local images are learned through the network.This thesis proposes a feature point detection and description algorithm based on deep learning.It uses a self-supervised method and self-label the feature point positions.The siamese network structure is used to construct matching point pairs to help the learning of descriptors.In addition,this thesis uses the location information of the feature points to enhance the descriptors,which can improve the distinguishability of the descriptor and reduce the intermediate storage of the network output.In order to make the deep feature suitable for real-time environment,this thesis designs a loss function,uses the constraints of the regular term to binarize the descriptor.While maintaining higher accuracy,it improves the matching speed of feature points.In this thesis,Hpatches datasets are used for experimental verification.The experimental verification shows that features and descriptors based on deep learning are better in performance than traditional handcrafted features.