Binocular Image Depth Information Extraction Based On SIFT

Binocular image depth extraction belongs to compute vision.The purpose of this paper is to imporve some existing algorithms for binocular image depth extraction.In the local algorithm,use SIFT feature and Absolute Intensity Difference(AD)as matching cost to speed up matching with keeping effect.In the global algorithm,use image segmentation to improve the effect of belief propagation.Specific research and imporvement are as follows:1.Research the principle of extraction depth information by binocular image,and the methods of matching cost measure,constraint conditions that make ill-posed stereo matching to be solvable in theory.On this basis,research matching cost aggregation method based on window,such as fixed window and multiple window algorithm.According to the thought of weighted template in image processing achieve an imporvement of fixed window.This imporvements effect is closed to multiple window,but its computational quantity only is 1/9 of multiple window.2.Using SIFT feature extraction to speed up line propagation algorithm.Line propagation is a local algorithm based on color similarity to segment image's scanline,and use line regions to aggregate matching cost.The matching cost measure method of original algorithm is the combination of AD and Census transform.Because the Census transform involves a lot of logic comparison,the computing time is too long In view of this,this paper uses SIFT feature extraction to replace Census transform to improve original algorithm.The combination of AD and SIFT feature is closed to the original algorithm in effect,but the computing time is reduced by nearly 75%.3.Using image segmentation to imporve belief propagation's effect.Belief propagation is an iterative algorithm based on Markov Random Field to minimize global energy function.When iteration number is large,the disparity is smooth over,resulting in effect of disparity discontinuities region worse.This paper use watershed segmentation and Canny edge extraction to stop propagation in image edges so avoid.smoothing over.After 20 iterations,the original algorithm,combining watershed and Canny's error matching rate is 4.79%,3.86%,3.25%,so the effect is improved.