A Research For Stereo Matching Of Binocular Vision Based On Visual Saliency And Machine Learning

Binocular stereo vision could create a depth map through two cameras.It is a key component in the field of computer vision.Binocular stereo vision is widely used in the fields of robot vision,automatic production,automatic navigation,VR,3D and so on.The stereo matching algorithm is the coretechnology of binocular stereo vision.The results of matching of the left and right images would directly affectthe accuracy of the 3Dreconstruction.Traditional algorithms mostly usethe objective data of image to complete the stereo matchingdirectly.Although this achieves a better matching result,it could be improved.According to the visual mechanism of the human,this study combined the saliency model and the stereo matching algorithm to improve the stereo matching of binocular vision.The main work is listed as follows:1.Traditional stereo matching algorithms,such as SAD,SSD etc.,only consider the objective characteristics of the pixel and ignore the human visual characteristics.Therefore there is a certain difference between the obtained disparity and the real subjective feelings of the human eyes.Based on the original algorithm of term weighting,this paper proposed a new ratio measurement method to describe the matching cost metric.Weights allocations were redefined after considering the RGB,grayscale,position,as well as the vision saliency features.The results showed a better subjective effect and less matching error rate.The algorithm has greater performance than the original algorithm.2.Traditional stereo matching algorithms usually treat a picture as the reference and the other as the image used for matching.There would be some block in some parts of images,so matching errors exit in the created disparity map.In this study,a method of image fusion was employed to solve occlusion problem.It individually used the two images as the reference map to calculate the parallax,and then it fusedthem to get the final parallax.Experiments showed that this method could solve the problem of occlusion to some extent,and couldadaptivelycalculate the edge and detail parallax.3.Traditional stereo matching algorithms are inefficient in per-pixel matching and calculating and cannot process large amounts of data in time.This study considered the idea of artificial neural network,so ELM training model was adopted to improve the efficiency.Firstly,the reduced left and right images were used to create the parallaxas training samples.Secondly,some pixel values and the corresponding parallax were chosen from the first step as training sample to train ELM model.Finally,we used the images to predict the parallax value and restore the disparity map.Experiments showed that this method can achieve a higher degree of efficiency while ensuring a certain matching accuracy.