Perspective template matching by differential evolution introducing the group remaking

Abstract

In order to achieve correct perspective template matching on two-dimensional general images, we propose a differential evolution (DE) introducing the group remaking. Since the perspective template matching has eight degree of freedom (DoF), acquisition of a global optimum, which is an area corresponding to a template in a target image, within limited calculation resource is difficult. In order to address this task, DE is utilized. This method can optimize some parameters simultaneously and acquire a good approximate solution within the limited calculation resources. The adopted DE divides the population into two equal groups, exploitation group for local search and exploration group for global search. By remaking the exploitation group every time, the best solution is found and updated, more effective search is possible. Also, linear group size changing scheme and initialization of parameters for crossover and mutation are introduced to acquire good performance. The experimental results using a general 2D image dataset show the lowest average error of the proposed method.

Performance Comparison of Population‐Based Meta‐Heuristic Algorithms in Affine Template Matching

“…To acquire a good result in affine template matching, population diversity is important [14]. This is because high diversity can prevent convergence of the population to local optima.…”

“…If f c is larger than f g,best , the best individual and its fitness are updated (lines 10-12). Otherwise, tc i is increased (lines [13][14]. Because the search by the employed bees use individuals, which are chosen by uniform random numbers, this search is global.…”

Artificial Bee Colony for Affine and Perspective Template Matching

“…Hence, checking all solutions is impractical because the computation cost becomes significantly high. To address this issue, evolutionary computation-based methods have been proposed [1][2][3][4]. In population-based meta-heuristic algorithms, candidate solutions (individuals) are randomly generated in a search space.…”

“…Although a global optimal solution cannot be obtained, this approach can acquire a practical solution with limited resource. Specifically, the differential evolution-based method achieves the best performance in affine (6 DoF) and perspective (8 DoF) template matching [4]. The primary objective of this study is to develop a method that can achieve the best performance.…”

“…Hence, a small the number of hyperparameters are desirable. Based on these factors, we focus on ABC to achieve better performance than the latest method [4]. In addition, the removal of the hyperparameter for a scout bee was addressed.…”