Blur invariant template matching using projection onto convex sets

Datum konání: 17.05.2019
Přednášející: Matěj Lébl
Odpovědná osoba: Kotera

Blur is a common phenomenon in image acquisition that negatively influences recognition rate if blurred images are used as a query in template matching. Various blur-invariant features and measures were proposed in the literature, yet they are often derived under conditions that are difficult to satisfy in practice, for example, images with zero background or periodically repeating images and classes of blur that are closed under convolution. We propose a novel blur-invariant distance that puts no limitation on images and is invariant to any kind of blur as long as the blur has limited support, non-zero values and sums up to one. A template matching algorithm is then derived based on the blur-invariant distance, which projects query images on convex sets constructed around template images. The proposed method is easy to implement, it is robust to noise and blur size, and outperforms other competitors in this area.