Research And Application Of Fast Image Matching Algorithm Based On Shape Gradient

With the development of automation in the manufacturing industry,machine vision technology has gradually replaced the traditional mechanical positioning technology by virtue of its non-contact,low cost and high efficiency.Image matching is indispensable for the manufacture,measurement and classification of high-quality products completed in the production line.Shape matching,as the most suitable technology for the recognition and positioning of workpiece in image matching,has always been a key problem in the research of machine vision.In this paper,a matching algorithm based on the gradient of image shape contour points is proposed to solve the real-time workpiece positioning problem in industrial production.In order to solve the problem of low search efficiency of traditional shape matching algorithm,an improved pyramid hierarchical search algorithm is proposed.The improved search strategy combines four methods of premature termination,midmatch termination,edge point sparsity and layer-by-layer overlapping filtering to reduce the computational burden of the algorithm.The edge matrix is convolved with the full 1 matrix to obtain the edge point distribution,and the regions with a small number of edge points are screened.The termination in the matching process is to terminate the matching process in advance according to the similarity measure.Edge point sparsity reduces the amount of template matching by equidistant sampling edge points.Layer-by-layer overlapping filtering eliminates invalid regions by nonmaximum filtering of candidate objects in each layer.The parallel algorithm uses SSE instructions to optimize the interpolation method of the gradient of sub-pixel point,and uses PPL parallel library to realize the parallel matching of multiple templates.Aiming at the problem of low precision of pose matching in traditional shape matching algorithm,a pose approximation algorithm combining small batch gradient descent method is proposed.This method uses the Canny algorithm based on quadric surface fitting to obtain the edge points of sub-pixel accuracy and improves the accuracy of the shape template.By minimizing the distance between the edge point and the corresponding edge tangent,the pose approximation problem is transformed into a nonlinear least squares problem to obtain more accurate pose parameters.In the process of iterative solution,a small batch gradient descent method is proposed to solve the pose approximation problem,which solves the problem of slow convergence rate.Aiming at the problem that the static shape matching algorithm cannot meet the needs of dynamic monitoring,a particle filter tracking algorithm based on shape gradient is proposed.The method uses second order autoregressive model to simulate the particle state transition,using Monte Carlo Simulation to approximate the probability density function of importance.The particle weight is updated by shape gradient similarity measurement,and the estimated position of the target is obtained by weighted average of particle state after resampling,which effectively reduces the search space of shape matching.Finally,in order to validate the performance of the improved matching algorithm,a shape matching debugging software is developed,and a series of comparative experiments are designed to test the improved algorithm in each part.By comparing with Halcon's shape matching algorithm and gray matching algorithm,it is proved that the improved algorithm has higher matching accuracy and anti-interference ability,and is obviously superior to other algorithms in running speed.