強化学習を用いた射影画像に対する画像照合

概要

論文の詳細を見る
The template matching technology is basic and important in robot vision, or in the field where a robot's activity is expected. This technology is very useful in the case that images are transformed in parallel or are slightly rotated. However, it can't be applied to the image with much scale conversion or projection-distortion, especially. In this paper, first, it is confirmed that even in the case that the scale conversion or projection-distortion is unknown, the target mark in the image is determined by the method based on four correspondent points between the template and target images. But, it is not easy to choose these arbitrary four characteristic points. So, next, the method, which finds the image appropriately and automatically for the template matching without using four characteristic points to give the projective transformation, is proposed. This method searches the optimum image for the template matching based on reward and value by using the reinforcement learning which has the same logic as human who moves to the appropriate place to estimate what the image means when being difficult to understand it.
2006-12-20