Research On Binary Feature Detection And Matching Algorithm In The Joint Light Field Spatial And Angular Domain

Image feature detection and matching are still important to many applications in computer vision.Traditional feature extraction and matching algorithms mostly focus on two-dimensional images.Two-dimensional images only record the position information of light rays and lose the intensity and scene depth information of the light rays.Due to the influence of illumination variation,imaging noise and geometric transformation,feature extraction and matching based on traditional two-dimensional images still have the problem of polysemy.Unlike traditional imaging,light field imaging technology combines computer vision with computational photography theory,breaking the bottleneck that traditional imaging can only record the light intensity distribution on the imaging plane.Light field imaging data simultaneously captures both the angular and spatial information of light rays,recording the light direction and intensity for each pixel.The angular sampling in light field data contains depth information of the scene and provides richer visual cues.Therefore,the structured multi-view data of light field cameras offer a new data foundation for solving the ambiguity problem in computer vision matching.The corresponding research work in this thesis includes:(1)In order to solve the limitations of two-dimensional image feature matching and the problem of multiplicity,this thesis proposes a light field spatial-angular domain joint binarized feature detection and matching algorithm.The method introduces the effect of feature point variation with viewpoints,and calculates the optimal view-invariant direction in the light field spatial-angular domain.For the selected subset of light field samples,a novel feature concatenation descriptor based on multi-scale block local binary patterns is introduced,namely the light field spatial-angular domain joint binarized feature descriptor.The proposed method calculates the cosine similarity of vector angles to determine the matching relationship between light field features.Experimental results demonstrate that the proposed features show superior matching precision and robustness,thus validating the advancement of the light field spatialangular domain joint binarized features.(2)To address the problem of low recognition accuracy of traditional binary pattern features in facial expression recognition tasks,this thesis proposes a method to enhance the robustness and precision of facial feature recognition by incorporating the multi-view information of the light field spatial-angular domain.Firstly,data preprocessing is conducted on facial expression datasets.Then,the spatial-angular domain joint multi-scale block local binary pattern feature descriptor is combined with facial expression recognition to extract vectorized descriptions of facial features.Finally,support vector machines are used to learn the feature description vectors for achieving high-precision recognition of different facial expressions.(3)Furthermore,based on the above theoretical analysis and experiments,this thesis designs a visualization system for the application of light field image features,including two functions of natural scene feature matching and facial expression recognition.The system is developed with web frontend technologies,the SpringBoot framework,and classic three-tier architecture,providing excellent user interaction capabilities.The completeness of the system’s functionalities is ensured through sample design and testing,aiming to assist users in conveniently applying the proposed methods for image feature matching and facial expression recognition.In conclusion,in order to effectively improve the precision and robustness of image feature detection and matching,this thesis proposes a joint multi-scale blocked local binary pattern feature extraction and matching algorithm on the spatial-angular domain.Through experiments in both real and virtual scenes on the paired light fields dataset,it is effectively verified that the precision of the multi-scale blocked local binary pattern features method proposed on the spatialangular domain in this thesis can reach 93%.And its matching precision is significantly better than the existing classical SIFT features and existing advanced light field features.Due to the borrowed idea of local binary features,the proposed features in this thesis also significantly outperform SIFT features and LIFF features in terms of quantization length,and also demonstrate good recognition performance in facial expression recognition experiments on the light field face dataset with support vector machine training classification.