Stereo matching is one of the key research directions in the field of computer vision,its core idea is to obtain the depth information of two images of the same object under different perspectives,and the obtained depth information is widely used in many tasks such as 3D reconstruction,robot navigation,and video surveillance,etc.Due to the stereo matching has the advantages of high reliability and low cost,it is usually used as an important way to obtain depth information in images.However,in order to obtain higher matching accuracy and computational efficiency,the stereo matching algorithms still need to face many unfavorable factors such as ill-illuminations,over-or under-exposures,depth discontinuities,weakly textured areas,and occlusion etc.In order to solve such problems,this paper designs several effective stereo matching algorithms based on sparse representation and deconvolutional network.The main research work is as follows:1.An unsupervised stereo matching cost based on sparse representation is proposed,which is independent of the ground truth disparity maps and has good robustness.Because supervised stereo matching costs need to learn model parameters from public datasets with the ground truth disparity maps,it is difficult to obtain ground truth disparity maps due to the high cost of ground truth disparity maps acquisition(which requires a lot of manpower and expensive instrument resources)and manually created ground truth disparity maps are prone to pixel-level annotation errors.Therefore,we design an unsupervised stereo matching cost that is minimally dependent on the ground truth disparity maps.In addition,stereo images are susceptible to external factors such as ill-illuminations,over-or under-exposures,etc.To reduce these effects,an unsupervised stereo matching cost based on sparse representation is proposed,which has a better robustness.In order to improve the computational efficiency of solving sparse representation coefficients,a parallel method is proposed,and its low computational complexity is verified by analysis.Experimental results show that the proposed unsupervised stereo matching cost based on sparse representation can obtain higher matching accuracy without relying on the ground truth disparity maps,and is not easily affected by the external factors such as ill-illuminations,over-or under-exposures.2.An unsupervised stereo matching cost based on two-branch convolutional sparse representation is constructed,which can reduce the redundancy of the convolution kernels,thus improving the computational efficiency.In order to reduce the computational complexity of the convolutional sparse representation,this paper proposes a novel two-branch convolutional sparse representation model,which can effectively characterize the geometric features of stereo image pairs with fewer convolution kernels,thus reducing the redundancy of convolution kernels and improving the computational efficiency.In order to solve the two-branch convolutional sparse representation model,an effective iterative algorithm is designed,and the algorithm has good convergence and low computational complexity.Then,the two-branch convolutional sparse representation model is used to construct an unsupervised stereo matching cost.The experimental results show that the unsupervised stereo matching cost based on two-branch convolutional sparse representation can not only improve the computational efficiency but also obtain higher matching accuracy.3.The stereo matching algorithm based on two-branch deconvolutional network is designed,which can capture deep features,improve the matching accuracy and reduce the computational complexity of the deconvolutional networks(DNs).To solve the two-branch deconvolutional network(TBDN)model,an iterative algorithm is designed,and related convergence analysis and computational complexity analysis are given to ensure effectiveness.Finally,based on the two-branch deconvolution network model,a disparity estimation network(DEN)structure is constructed,which uses the weights obtained by the two-branch deconvolutional network as the initial weights.The experimental results show that the proposed stereo matching based on the two-branch deconvolutional network(TBDN)can improve the computational efficiency,and the matching accuracy of the proposed stereo matching algorithm based on two-branch deconvolutional network is higher than that of other similar methods. |