A Study On Mid-Level-Feature-Based Template Matching Algorithms

Template Matching is a well-known and active research topic in image processing and computer vision fields,and it has been widely used in industrial manufacture,medical services and military,and so on.After a template image and a scene image are given,template matching method is employed to find the most similar candidate window in the scene image as the matched regions.To meet practical vision-based position requirements in electronic manufacturing equipments,for instance robot pick up and counting object,template matching method should have high robustness,speed and accuracy.However,most of template-matching methods can not deal with those challenges well.Thus,in this study,three template methods based on the mid-level feature are proposed to deal with those challenges in the electronic manufacture.The detailed works are as follows:(1)A novel gradient direction code template matching(GDC-TM)algorithm is proposed to locate the sharp-edge objects with high matching speed and matching accuracy.In the algorithm,the gradient direction of edge point is coded as 8-bit code,which is used in template compression for less of information,and is used to calculate the similarity fast.Besides,a very high accurate matching result is obtained by the approaching method between the point pairs in fit lines of edge points.The average runtime of GDC-TM is 10 ms and the matching accuracy is ±0.06 pixels with the image size 512×512pixels;(2)A novel superpixels-based fast normalized cross correlation(SPNCC)algorithm is proposed to locate an object in blur case with high matching speed.In the algorithm,the superpixels are utilized to compress the template not evenly,and an adaptive step search strategy is used to make matching faster.The algorithm is robust against blur,noise and non-linear illumination.SPNCC is 2 times faster than AMW method which was the fastest NCC method;(3)A novel robust semantic template matching(RSTM)algorithm is proposed to locate a texture object with 8 different interferences.In the algorithm,a modified SLIC algorithm,KD-SLIC,is used to obtain regular and boundary-adhere-well superpixels,which the superpixel region binary descriptors are constructed based on.Besides,the marginal probability model is applied.Comparing with other 7 methods,RSTM has the highest IOU value,0.984.In the study,some application tests in electronic manufacturing equipments are introduced.The stencil printing accuracy is ±0.015 mm using GDC-TM in the GKG project.In other tests about electron components and chips,the recognition rates of SPNCC and RSTM are not less than 99.2%.