Research On Template Matching Algorithm For Deformation Image

Template matching is a method of registering two images in spatial context.As a key technology,it has been widely used in many fields such as autonomous target recognition and tracking.Since the fact that the images often contain various deformations in actual matching scenes,and the traditional matching methods usually adopts the strategy of ergodic searching by per-pixel,the matching algorithms often fails to meet the demands of both precision and speed.Therefore,this thesis probes into the accuracy and rapidity of the deformation image matching algorithms under various scenarios and carries out the following research.In order to solve the problems that the random searching method is easy to fall into the local extremum and the two-dimensional affine transformation space traversal searching method is inefficient,this thesis puts forward the Mutation Particle Optimization Fast Matching Algorithm(abbreviated as MPO)featured with automatic optimization.Based on the affine transformation model,MPO combines the randomness and mutation mechanism to improve the global optimization ability,and adopts the discrete sampling method to reduce the computation so as to improve the matching speed.Besides,in view of the deviation of the attitude parameters provided by the actual imaging platform,this thesis puts forward the Perspective Transformation Matching Algorithm with Biased Attitude Parameters(abbreviated as BPT).Based on the prespective transformation model,BPT uses the same idea of automatic optimization in MPO to correct the biased parameters.Experiments results show that the MPO has better matching performance,and that the BPT can eliminate the attitude deviation in the iterative process effectively.The results of software simulation also proves that the two algorithms can meet the requirement of real-time engineering which can be applied to the realistic matching scene.In order to improve the insufficient matching ability of the parametric template matching algorithms for complex deformation images,this thesis proposes the Template Matching Algorithm Based on Deconvolution Network(abbreviated as DM-Net)on the basis of deeping learning theory,which further expands the applicability of template matching algorithms.Combining with the characteristics of siamese network and 2ch network,DM-Net improves matching precision through improving the network structure and loss function,and improves matching speed through adding the deconvolution network for image coding.In addition,this thesis also analyzes the influences of each structural alteration on the network performance,tests the matching results of the network on the standard dataset and the image sequence of imaging platform,and compares it with other network models afterwards.The experiment results show that DM-Net has strong sensitivity of positive sample and the resolving ability of negative sample.