Research On Feature Detection And Matching Algorithm Of Light Field Image Based On EPI

Image feature detection and matching is a basic research problem in the field of computer vision.At present,although the traditional algorithms are in the mature stage of theory,they also face a series of challenges,such as occlusion,shadow,similar texture and illumination changes in complex natural scenes.One of the fundamental reasons is the low computability of traditional image data.In order to break through the bottleneck of low computability of two-dimensional image,light field imaging technology uses four-dimensional parametric method to obtain the position and angle signals of spatial light.Based on the theory of light field imaging,a feature detection,description and matching algorithm of light field epipolar plane image light field is proposed in this paper.Different from the traditional methods,this paper constructs the gradient direction histogram in the light field epipolar-plane image to detect the peak distribution,and describes the feature points from the spatial domain and the angle domain respectively.The main research work of this paper is as follows:(1)This paper analyzes the epipolar plane image space of the light field,constructs the gradient direction histogram in the 3*3 neighborhood of the pixel,and extracts the key points with larger peaks.According to the contrast between the key points and other pixels in the neighborhood,the stable key points with greater contrast are retained.(2)This paper proposes a 3D local feature descriptor.Based on HoG feature description algorithm,this paper uses local neighborhood of feature points to construct 164 dimensional feature descriptors from horizontal angle domain,vertical angle domain and spatial domain.In order to enhance the performance of feature descriptors,a scale invariant description vector is added to the spatial domain to cope with the scale changes of features.(3)This paper proposes a feature matching method based on weighted similarity measure.The similarity of feature vectors is calculated by cosine of vector angle,and the weights of feature vectors in horizontal angle domain,vertical angle domain and spatial domain are allocated to improve the accuracy of feature matching.Compared with the traditional methods SIFT,HoG and two existing light field feature detection algorithms,the experimental results on real and virtual scene light field data show that the proposed light field epipolar plane feature detection,description and matching algorithm can significantly improve the repetition rate of feature detection and the accuracy of feature matching.