空间约束下异源图像误匹配特征点剔除算法

Algorithm for Eliminating Mismatched Feature Points in Heterogeneous Images Pairs Under Spatial Constraints

红外与可见光图像因其显著的光谱特性差异,在配准过程中易出现特征点误匹配率高的问题。当前广泛应用的误匹配剔除算法通常采用随机采样结合模型拟合的策略,这类方法往往难以兼顾配准精度和速度,表现为算法迭代次数过高或鲁棒性不强。针对这一问题,提出一种基于空间约束的优先采样一致性(SC-PRISAC)误匹配剔除算法。利用材料辐射率差异设计兼具红外与可见光特征的双光谱标定靶标,基于双边滤波金字塔标定获取相机内外参数,在此基础上利用极线约束定理和深度一致性原则构建异源图像间的空间约束关系。使用高质量特征点优先采样策略减少了算法的迭代次数,有效剔除误匹配特征点。实验表明:所提算法实现了亚像素红外与可见光双目标定,标定误差降低至0.430 pixel;在提高配准精度的同时,也有效提升了处理速度,单应性矩阵估计误差为7.857,处理时间仅为1.919 ms,各项性能均优于RANSAC(random sample consensus)等算法。所提算法为红外与可见光图像配准提供一种更为可靠和高效的误匹配剔除解决方案。

Objective Infrared and visible light images exhibit significant differences in spectral properties due to their distinct imaging mechanisms.These differences often result in a high mismatch rate of feature points between the two types of images.Currently,widely used mismatch rejection algorithms,such as random sample consensus(RANSAC)and its variants,typically employ a strategy of random sampling combined with iterative optimization modeling for consistency fitting.However,when aligning heterogeneous images with high outlier rates,these methods often struggle to balance alignment accuracy and speed,leading to a high number of iterations or weak robustness.To address the relatively fixed positions of infrared and visible detectors in dualmodal imaging systems,we propose a spatial constraints priority sampling consensus(SCPRISAC)algorithm.This algorithm leverages image space constraints to provide a robust inlier screening mechanism and an efficient sampling strategy,thus offering stable and reliable support for the fusion of infrared and visible image information.Methods In this study,a bispectral calibration target with both infrared and visible features is designed based on differences in material radiance.We achieve highprecision binocular camera calibration by accurately determining the internal and external parameters of the camera using a bilateral filtering pyramid.Based on this calibration,the spatial relationship between heterogeneous images is constructed using the epipolar constraint theorem and the principle of depth consistency.By implementing a priority sampling strategy based on the matching quality ranking of feature points,the number of iterations required by the algorithm is significantly reduced,allowing for precise and efficient elimination of mismatched feature points.Results and Discussions Our method’s calibration accuracy is assessed through the mean reprojection error(MRE),with comparative results presented in Table 1 and Fig.7.The findings demonstrate a 58.2%improvement in calibration precision over the spot detection calibration technique provided by OpenCV,reducing the calibration error to 0.430 pixels.In the outlier rejection experiment,the progression of feature point matching across stages is detailed in Table 2.Following the introduction of spatial constraints,all valid matches are retained,and 27 outlier pairs are discarded.An additional 10 outlier pairs are further eliminated through preferential sampling strategies.To comprehensively evaluate the algorithm’s performance,several comparative methods,including RANSAC,degenerate sample consensus(DEGENSAC),MAGSAC++,graphcut RANSAC(GCRANSAC),Bayesian network for adaptive sample consensus(BANSAC),and a neural networkbased∇-RANSAC,are employed,with evaluations based on inlier counts,homography estimation errors,accuracy,and computational runtime as shown in Table 3 and Fig.12.The proposed algorithm achieves a notably low homography estimation error of 7.857 with a runtime of just 1.919 ms,outperforming all comparative methods.This superior performance is primarily due to the SCPRISAC algorithm’s robust spatial constraint mechanism,which effectively filters out outliers that contradict imaging principles,enabling more accurate sampling and fitting.In addition,the robustness of the proposed method and competing algorithms under complex scenarios is investigated by varying the proportion of outliers in initial datasets,as illustrated in Fig.13.All algorithms perform satisfactorily when outlier ratios are below 45%.However,as the outlier ratio escalates,the precision of traditional methods like RANSAC deteriorates significantly.Remarkably,even at an extreme outlier ratio of 95%,SCPRISAC maintains an accuracy rate of 70.2%,whereas other algorithms’accuracies drop to between 12%and 49%.These results highlight the significant advantage of the proposed method in scenarios with high mismatch rates,demonstrating its superior applicability and effectiveness in aligning infrared and visible light images under challenging conditions.Conclusions To address the challenge of high mismatch rates in infrared and visible image alignment,we propose an algorithm for rejecting mismatched feature points based on optimizing camera spatial relations.By designing a bispectral calibration target and improving the circular centroid positioning algorithm,subpixellevel infrared and visible binocular camera calibration is achieved,with the calibration error controlled within 0.430 pixel,significantly enhancing camera calibration accuracy.The algorithm integrates spatial constraints based on epipolar geometry and depth consistency to accurately exclude mismatched features that violate physical imaging laws and reduces computational complexity through an intelligent sampling strategy that prioritizes highquality feature points.Experimental results show that the proposed method achieves a homography estimation error of 7.857,a processing speed of 1.919 ms,and maintains excellent performance even under high outlier ratios,outperforming other mismatched feature point rejection algorithms and proving its superior generalization and reliability in addressing infrared and visible image alignment problems.