Image features invariant with respect to blur

Abstract The paper is devoted to the feature-based description of blurred images acquired by a linear shift-invariant imaging system. The proposed features are invariant with respect to blur (this means with respect to the system point spread function), are based on image moments and are calculated directly from the blurred image. This way, we are able to describe the original image without the PSF identification and image restoration. In many applications (such as in image recognition from a database) our approach is much more effective than the traditional “blind-restoration” one. The derivation of the blur invariants is the major theoretical result of the paper. Several experiments are presented to illustrate the efficiency of the invariants for blurred image description. Stability of the invariants with respect to additive random noise and boundary effect is also discussed and is shown to be sufficiently high.