Research On Object Matching And Detection Based On Spatial Consistency

Object matching is a fundamental problem in the field of computer vision and pat-tern recognition.Its purpose is to find different instances of a object from two or more images and establish a correspondence between them.Many applications such as image recognition,image inpaintings,object tracking,3D reconstruction,and action detec-tion,etc.benefit from accurate object matching.The standard object matching problem contains two key steps:feature matching and spatial consistency measurement.They are either done separately or simultaneously.Spatial consistency measurement requires establishing a mapping model between the objects and predicting the model parame-ters according to the image features.Although the performance of feature matching can be improved using the measurement of spatial consistency,the performance improve-ment is limited by traditional measurement methods,due to the influence of illumination changes,noise,outliers,textureless and object deformation in natural images.Therefore,studying efficient and robust object consistency computing methods has significant the-oretic and practical values.This paper researches the object matching based on spatial consistency between an image pair and within an image.For the latter,this paper focuses on a more complex case that is repetitive object detection.This paper uses model correla-tion to optimize the linear matching algorithm.This algorithm can discover the optimal transformation model from the large scale space to achieve object matching between an image pair.For the repetitive object detection,we establish a match propagation algo-rithm based on pattern alignment to solve the multiple object matching problem.The main works of the thesis are summarized as follows:First,we propose a residual-consensus driven linear matching?RDLM?algorithm for pairwise images.In the object matching system,the problem of simultaneously solving the geometric transformation model and the point correspondence is non-linear.Because linear programming?LM?is a simple and effective method for modeling ob-ject matching and can handle large-scale matching problems,we convert the nonlin-ear quadratic matching problem into a LM problem.Existing LM methods usually use low-order transformations to artificially initialize a linear model.The proposed RDLM algorithm generalizes existing LM methods by allowing higher-order transformations to automatically initialize a linear model.Based on the observation that transforma-tion models generated from inlier subsets exhibit correlated behaviors?termed”residual consensus”hereafter?,we develop a residual-consensus robust estimation algorithm to project the non-trivial linear transformation problem into a much smaller subspace,and thus enable efficient optimizations through linear programming.Experiment results on synthetic and real databases demonstrate the effectiveness and robustness of the pro-posed algorithm.Second,we propose a coupled detection and segmentation algorithm for repeti-tive patterns based on match propagation?CpdDSR?in real-world images.Repetitive objects,also known as repetitive patterns,are high-level image structures that exist ubiq-uitously in natural and man-made environments.Automatically and robustly detecting various types of such patterns from an real-world image is still a challenging problem.Compared with pairwise image matching,repetitive pattern detection needs to solve a series of problems such as textureless,strong noise sensitivity,multiple object de-formations,and unknown template.The proposed CpdDSR algorithm can solve these problems well.This algorithm optimizes the multi-layer joint alignment approach with an iterative growing strategy so as to achieve spatial consistency among patterns.The objective of alignment is to minimize the photometric and geometric variations mod-eled by the Lie group and Gaussian mixtures.In multi-layer segmentation,we employ a region-growing strategy to iteratively discover the complete repetitive structure.First step is to detect deformations of a pattern using a pattern matching technique.Next,a multi-layer joint alignment is applied to the detected patches to improve the matching performance.Extensive experiments and comparative analysis in PSU-NRT and Sym-metry Detection data set demonstrate the effectiveness of the proposed algorithms for the detection and segmentation of repetitive patterns.Finally,on the basis of the proposed coupled detection and segmentation frame-work for repetitive pattern based on match propagation,regarding the advantages of the functional map model,we study repetitive pattern alignment by embedding segmenta-tion cue with a functional map model.A consistency functional map propagation al-gorithm?CFMP?is proposed,which extends the functional map model to incorporate dynamic propagation.Hence,the pattern detection is only performed in a local region during each expansion,minimizing the effects of photometric and geometric variations.The establishment of a propagation model requires two-step iterative operations?The first one aligns the patterns from a local region,transferring segmentation functions among patterns.It can be cast as an L_2,1norm optimization problem.The latter step up-dates the template segmentation for the next round of pattern discovery by merging the transferred segmentation functions.Experimental results demonstrate that the proposed CFMP algorithm works substantially better than other state-of-the-art repetitive pattern detection methods in PSU-NRT and Symmetry Detection data set.Although the run time of the CFMP algorithm is slightly lower than the CpdDSR algorithm proposed in this thesis,the CFMP algorithm achieves superior detection performance w.r.t.CpdDSR algorithm.So the CFMP algorithm is suitable for computer vision systems with higher precision requirements.