A Study On Hierarchical-Semantic-Feature-Based Image Matching Algorithms

Intelligent manufacturing is a strategic development trend to improve the level of industrial production,and machine vision is the key technology to realize intelligent manufacturing.In the field of image processing,template matching is an important theoretical basis of machine vision,which is widely used in visual positioning,robot grasping,target tracking,3D reconstruction et al.The complexity of matching scenarios and the increasing production standards of industrial production require the template matching algorithms to have high matching speed,accuracy and robustness.To meet the requirements of vision positioning in different application scenarios and overcome the shortcomings of the existing template matching algorithms which only utilize the features of single layer and low layer for template matching,in this paper,three template matching algorithms based on the hierarchical semantic features are proposed:(1)To meet the demand of rapid and accurate positioning of components with clear outlines in industrial structured scenarios,a novel template matching algorithm based on hierarchical gradient features(HGF)is prposed in this study.The algorithm utilizes gradient information to express the object features,contour features and detail features of a target.Then,the hierarchical search strategy based on the Latent Dirichlet Allocation model is proposed,which improves the matching efficiency of the template matching on the premise of ensuring the matching accuracy.Compared with existing algorithms such as SBM and GHT based on gradient features,the matching speed is improved by 3～4 times.(2)To meet the requirement of robust positioning of targets with rich texture information for intelligent robots in natural scenarios where multiple disturbances exist simultaneously,a robust hierarchical semantic part features based tree matching method(RHSP)is proposed.Based on hierarchical semantic parts,the algorithm expresses the template image and the scene image into a hierarchical semantic part feature tree and a hierarchical semantic part feature forest respectively,which improves the robustness of the template expression.RHSP models the template matching task as a tree matching problem.In order to execute tree matching,a tree matching framework based on bayesian network model is proposed to improve the robustness of the matching process.RHSP algorithm is compared with classic NCC,MTM,SURF and the latest dsp-sift,TFeat,Hard Net and other template matching algorithms on the COCO dataset,and the experimental results show that RHSP algorithm achieves the highest precision 0.9795.(3)To meet the robust positioning requirements of targets with clear structure information in natural scenes under large deformation or large view transformation interference,a novel hierarchical semantic part features based graph matching method(HSPF)is proposed.HSPF utilizes the high-level semantic part features to rapidly position the most similar area to the template image in the scene image.And then HSPF utilizes middle-level semantic part features to express the template image and the candidate area into graph structures.Based on graph matching,the appearance of the middle-level semantic parts and the topology information between parts are simultaneously utilized to improve the robustness of template matching.HSPG-GM is compared with template matching algorithms such as NCC,MTM,SBM,GHT and SURF on COCO data set,and the results show that HSPG-GM algorithm achieves the highest precision of 0.978 under the interferences of deformation and view changes.The three template matching algorithms proposed in this paper are tested in their respective applicable application scenarios.The HGF algorithm achieves a 100% counting rate on the high-speed matching platform,and the average time of an image with 300*400 pixels was 0.2s(Matlab Platform).RHSP algorithm achieves an average positioning accuracy of 0.91 in 100 groups of application data sets containing both motion blur and geometric transform interference.The HSPF algorithm achieves matching rates of 99% and 100% respectively in the deformation and view change test sequences of 500 images.