Research On Robust And Efficient Template Matching In The Wild

Template matching is a fundamental and important research topic in computer vision.This technique has a wide range of applications in many fields,such as general object detection,target location in machine vision systems,object tracking in video surveillance systems,image stitching,fine-grained recognition of goods and so on.At present,most template matching algorithms mainly focus on solving the problem of gray-scale images in industrial scenes.However,the performances of these template matching methods have unsatisfied performances in the wild and far below the demand of application.Color images in the wild are quite different from grays-scale images in industrial scenes,which are usually more complex and variable.Therefore,the robust and efficient template matching method in natural scenes remains to be solved academically.Recently,a series of methods based on nearest neighbors field have appeared in the academic research.These methods greatly improve the performance of template matching in natural scenes.Nevertheless,they impose strong assumptions on data that make algorithm very sensitive to large rotations and large deformations and the computational complexity of them are so expensive that is hard to apply to real-world applications.In this paper,we would reveal the importance of global semantic information based on nearest neighbor field,and then propose an innovative novel global-aware diversity(GAD)similarity metric.The GAD similarity metric uses global semantic information to supervise local diversity information,enabling it to further filter out background and outliers to increase robustness.At the same time,in order to meet the runtime requirement in the real scene,this paper will elaborately design an efficient algorithm implementation.Specifically,the integral image is used to efficiently calculate the global semantic information,and the buffer is used to quickly update the local diversity information.As a result,the complexity of the algorithm is greatly reduced from O(|I| · |?|)to O(|I|)and runtime can be accelerated by 10~3 to 10~4times.In order to measure the robustness of the GAD similarity metric proposed in this paper and the efficiency of its corresponding algorithm implementation,this paper will compare the proposed GAD algorithm with the state-of-art algorithms on four challenging datasets and do full analysis to the results.Experiments will show that the GAD algorithm is superior to the previous methods with non-deformation loss both in accuracy and running time,and is robust than methods with deformation loss in case of large rotation or large deformation.In addition,with certain settings,the proposed GAD algorithm only takes 3ms to calculate similarity metric on high-resolution(1280×720)images,achieving near real-time performance.