ECN>> convolution-template-matching>> 返回
项目作者： sunsided

项目描述 ：
Template Matching by Convolution in MATLAB
高级语言： Matlab
项目地址: git://github.com/sunsided/convolution-template-matching.git
创建时间： 2016-08-18T00:22:24Z
项目社区：https://github.com/sunsided/convolution-template-matching
开源协议：
下载


Template Matching by Convolution

Template matching attempts to find instances of a given template in an existing image by finding areas of maximum correspondence. While this can be done in terms of a cross correlation, care has to be taken to normalize both input and template, as cross correlation by itself is not invariant to mean shifts.

This experiment follows the idea that convolution is the same operation as cross correlation, when all axes of the template (i.e. the kernel) have been flipped; in terms of a two-dimensional correlation with a template, this results in a two-dimensional convolution of the same template, rotated by 180°; so

  1. xcorr2(image, template)

is conceptually the same as

  1. conv2(image, fliplr(flipud(template)) % or
  2. conv2(image, rot90(template, 2))

Finding the template matches is then a question of finding the local minima (i.e. areas of maximum correspondence).