Stereo Matching Algorithm Base On The Improved Patch Matching Network

In human vision,it is a very convenient and critical function to use the eyes to obtain the depth of an object in the scene.Depth information plays a very important role in the period of high-speed development for computer.Stereo vision is one of the computer vision methods to simulate human vision.By using two synchronous cameras to simulate human eyes,stereo vision algorithm is used to obtain the depth of the object.Among the stereo vision steps,obtaining disparity by stereo matching is not only the most difficult,but also the most significant step in stereo vision.With the development of visual algorithms,stereo matching methods appear one after another.And in recent years,by the extraordinary performance in object detection,image synthesis and other image fields,convolutional neural network is widely used to solve stereo matching issues.How to design an appropriate network to achieve high-accuracy and high-efficiency calculation is still a difficult problem to be solved urgently.In this paper,based on the idea of patch matching network,by improving the feature extraction network structure,classification network structure and the loss function used in training convergence,a matching method with certain competitiveness in the accuracy and efficiency of disparity obtaining is proposed.The main contributions of this paper are as follows:1)Aiming at solving the time-consuming problem of traditional patch matching network,a reduction classification network structure is proposed.At the same time,the batch normalization layers are applied to relief the gradient vanishing problem and accelerate the convergence of the network.The improved network structure saves a lot of runtime and improves the accuracy of the disparity results.2)The residual network and feature extraction are combined to compensate the defect that the test accuracy decreases due to a simply deeper network.At the same time,we propose two improvement loss function based on cross entropy for the above two network respectively.The experiment results show that the improved network and loss function have better performance than the traditional patch matching network,and there is a significant decrease in time consumption.3)Before the improved deep learning method,we propose an energy function based on the traditional segmentation method.By supplementing gradient information and other optimization such as interpolation in the process of segmentation and fitting,the proposed method can also obtain a dense disparity.Compared with some other traditional methods,our method has a considerable performance.After improving the patch matching network structure and loss function,the proposed algorithm increases the running speed by about 32% compared with the previous matching method.At the same time,the subjective disparity result generated by the algorithm reduces the mismatch and error blocks compared to the classical algorithm,and can obtain a denser and more accurate disparity map.It is used in fields such as depth measurement and 3D reconstruction with higher accuracy.