Line Feature Description And Matching Of Multi-Angle Images

Feature matching is an indispensable technology in the field of computer vision and plays an import role in many applications.Line matching,a kind of feature matching,has attracted much attention because it contains the structural information of objects and has the capacity of resisting disturbance.In recent years,many scholars have done a lot of research on line matching.However,due to the inherent difficulties of line matching and the inaccurate location of line endpoints,line matching has made little progress.At present,the main algorithms of line matching are as follows:firstly,the feature lines are detected from the image,and then by using the neighborhood information of the feature line to construct line descriptor.Finally,the similarity of the line descriptor is calculated to determine whether the line matches and these methods have achieved effective results.However,the construction of line descriptors and the computational complexity of similarity measurement are relatively high.When the image has scale transformation or rotation,the line matching performance is reduced significantly.As a result,this paper proposed a novel line matching algorithm with high matching efficiency and good robustness to scale transformation and rotation.A novel line matching network named MobileNetV2-FPN is proposed to improve the efficiency of line matching and solve the scaling problem.MobileNetV2 using separable convolution,which can greatly reduce the complexity of the network,instead of the traditional convolution to extract line features.To make the line matching algorithm more efficiency,this paper proposed a new network based on MobileNetV2.We use MobileNetV2 to extract line features of different depths to make the extracted features contain different scale information,and then line matching is performed by concatenating these different scale information through classification layer,which can greatly improve the robustness of the network to scale transformation.To solve the rotation problem of line matching,this paper proposed a new dataset generation method.Rotate the detected line to the horizontal direction first,and then intercept the image block of fixed size in the horizontal direction with the straight line as the center.By this method,a pair of matching lines can always be intercepted to the same neighborhood without considering the angle between the straight line and the horizontal direction,which can improve the line matching performance of rotating images.This paper proposed a novel line matching algorithm based on deep convolutional neural networks.Compared with the line descriptor used in traditional methods,which can generally capture low-level line features,the proposed algorithm can extract not only low-level line features but also middle and high-level line features.Therefore,the matching accuracy can be greatly improved if the network can extract multi-level and stable line features.The proposed algorithm has been tested on three datasets collected under different conditions and compared with different algorithms.To analyze the robustness of our proposed algorithm,the three test datasets are processed differently,including rotation and scale transformation.We calculate the average time of processing images pairs in each dataset by different algorithms.Experimental results show that the proposed algorithm can achieve better matching performance.It is a technically feasible line matching algorithm.